How Pragmatics Shape Articulation: A Computational Case Study in STEM ASL Discourse

Alicia Wooten; Erin Walker; Laurel Aichler; Lee Kezar; Lorna Quandt; Malihe Alikhani; Mert Inan; Saki Imai

REVIEW 2 major objections 5 minor 20 references

How Pragmatics Shape Articulation: A Computational Case Study in STEM ASL Discourse

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ASL signs produced in dialogue are shorter and more compact than isolated signs, and repeated mentions compress further—an effect that disappears when the same signer signs alone.

desk verdict Descriptive kinematics are solid, but the entrainment-vs-effort distinction hinges on a single underpowered cross-modality null that needs reporting or softening. read the letter →

arxiv 2510.23842 v2 pith:KU7JNJD2 submitted 2025-10-27 cs.CL

Saki Imai , Lee Kezar , Laurel Aichler , Mert Inan , Erin Walker , Alicia Wooten , Lorna Quandt , Malihe Alikhani This is my paper

classification cs.CL

keywords AmericanSignLanguagephoneticreductionentrainmentmotioncaptureSTEMdiscoursedialoguemodelspragmatics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the physical form of a sign depends on whether the signer is talking with someone or signing alone. Using motion-capture recordings of an instructor–student biology dialogue, isolated vocabulary items, a solo lecture, and interpreted articles, the authors show that dialogue signs are 24.6% to 44.6% shorter in duration than isolated signs and grow progressively more compact with each repeated mention. Crucially, the same instructor shows no such reduction when signing alone, which the authors take as evidence that the compression is interactional entrainment with the interlocutor, not just an individual habit of speaking faster. The paper also shows that current sign-recognition models trained on isolated or interpreter-style data perform poorly on dialogue signing, implying the models miss the articulatory variability that characterizes real conversation. A sympathetic reader would care because it quantifies a pragmatic force in sign language and points to a concrete gap in sign-language technology.

What carries the argument

The argument is carried by continuous kinematic measurements of the hands, arms, and fingers—spatial extent (bounding-box diagonal), path length, average velocity, articulation duration, and vertical hand position—extracted from motion-capture recordings at 120 frames per second. The load-bearing comparison structure is the same-signer monologue control: by comparing the instructor's repeated productions of the same STEM signs in dialogue versus a solo lecture, the design separates interactional entrainment from individual effort reduction. A second machinery piece is the evaluation of two pre-trained sign embedding models on a continuous sign-spotting task, which tests whether models traine

What would settle it

Measure the same instructor's repeated STEM signs in a solo lecture using the same 3D motion-capture rig and a matched set of repeated tokens (e.g., all signs with at least three mentions), and test whether sign duration decreases with mention order; if a statistically significant negative slope appears (p<.05), the paper's central claim that dialogue compression arises from interactional entrainment rather than individual effort reduction is undermined.

Watch

Extended reading notes

Core claim

The central discovery is that ASL articulation is measurably shaped by dialogue pragmatics: signs in a two-person STEM conversation are spatially smaller and temporally shorter than the same signs produced in isolation, and compression increases across repeated mentions. Duration reductions average 24.6% for the instructor and 44.6% for the student relative to isolated productions, with significant duration-slope correlations across mentions in dialogue (instructor r=0.279, p<.001; student r=0.295, p<.001). In the same instructor's solo lecture, no significant duration change appears (r=0.026, p=0.895), which the authors read as evidence that dialogue reduction reflects entrainment between i

Load-bearing premise

The load-bearing premise is that the monologue control is a valid and sufficiently powered basis for concluding that reduction is dialogue-specific; the null duration result in monologue comes from a single same-instructor video recorded with a different (2D video-based) pose-estimation pipeline than the dialogue's 3D motion capture, and the paper does not report how many repeated tokens or signs entered that monologue test.

Editorial extensions

If this is right

If dialogue signing is systematically shorter and more compact than isolated signing, sign-language datasets built from isolated vocabulary or interpreter performances underrepresent natural articulation.
The absence of repetition-driven reduction in monologue implies that entrainment—mutual adaptation between interlocutors—is a measurable articulatory force in ASL, not just a lexical or acoustic phenomenon.
Current sign embedding models generalize poorly from monologue/isolated data to dialogue; practical sign-searching and educational tools built on such models will miss or misrank naturally produced signs.
The selective reduction of non-dominant-hand movement in dialogue suggests a targeted economization (weak drop) that preserves clarity while reducing redundant effort, a pattern that models would need to learn.
The instructor's asymmetric adaptation toward the student (embedding slopes consistently positive for instructor-to-student, negative for student-to-instructor) suggests role and power dynamics shape articulatory convergence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If this holds, a concrete testable prediction left implicit is that in a dialogue between two signers matched for status and prior familiarity, compression slopes should be symmetric; in an asymmetric teacher–student pair the higher-status signer should show shallower reduction, as the instructor does here.
A direct extension would be to record the same signer in monologue and dialogue with the same motion-capture system and matched token counts; if a significant monologue duration reduction then appears, the entrainment interpretation would need revision in favor of effort reduction.
If confirmed, this suggests sign-language technology evaluation should include a pragmatic-variance stress test—dialogue samples with repeated mentions—alongside standard isolated-sign benchmarks, since current monologue-style benchmarks may overstate real-world readiness.
The absence of significant sign lowering in dialogue, despite clear spatial and temporal reduction, hints that vertical position is governed by different constraints (prosody, coarticulation) than horizontal or temporal compactness; a follow-up could test whether sign lowering appears only in more casual or faster interactions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my take. The paper's real contribution is the data: a Vicon motion capture of a spontaneous ASL STEM dialogue between two deaf signers, with the same student's isolated productions and the same instructor's recorded lecture as controls. That allows a within-signer, same-modality comparison of dialogue versus isolated vocabulary—the student shows 44.6% shorter durations in dialogue, the instructor 24.6%—and a clean within-dialogue repeated-mention analysis showing significant compression for both. The finding that SignCLIP and I3D embeddings fail to retrieve these signs in dialogue is a useful, concrete demonstration of the domain gap. That part should be published.

The soft spot is exactly what the reader flagged: the monologue control that carries the entrainment conclusion. The null r=0.026, p=0.895 comes from one Atomic Hands video of the instructor, processed with MediaPipe 2D, while the dialogue data are Vicon 3D. No token counts, no confidence interval, and the excerpts were hand-selected for keyword overlap. That null cannot support the strong claim that reduction is absent in monologue; it may just be underpowered or attenuated by the tracking modality. The paper's own limitation section is honest about small N and hearing annotators, but it doesn't address this specific measurement mismatch.

Two other wobbles. The discussion calls the work 'the first quantitative evidence' for pragmatic adaptation in sign articulation; Hoetjes et al. 2014 already quantified sign reduction across repeated references, so the novelty claim needs qualifying. And the embedding-based entrainment metrics in §4.3 are built on encoders that perform at near-zero recall on this data; those results are suggestive at best.

On balance, this deserves a serious referee. The descriptive result—dialogue signing is spatially and temporally compressed relative to isolated vocabulary, with compression increasing over repeated mentions—holds up as a case study. The interpretation that this is interactional adaptation rather than generic effort reduction requires the authors to either strengthen the monologue control or soften the conclusion to 'consistent with entrainment.' I'd send it out, with the request to report power/token counts for the monologue null and to use per-signer baselines for the headline percentages.

Referee Report

2 major / 5 minor

Summary. The paper presents a motion-capture case study of an 8.52-minute ASL STEM dialogue between an instructor and a student, supplemented by isolated vocabulary productions from the student and monologic videos (a solo lecture by the same instructor and interpreted ASL STEM Wiki content). It computes kinematic features (spatial extent, path length, duration, velocity, vertical position) and compares (i) dialogue vs. isolated articulation, (ii) repeated mentions within dialogue vs. within monologue, and (iii) the ability of pretrained SignCLIP and I3D encoders to spot STEM signs across these contexts. The main empirical claims are that dialogue signs are 24.6%–44.6% shorter in duration than isolated signs, that repeated mentions in dialogue show spatial/temporal reduction that is absent in the monologue, and that current sign-embedding models do not generalize to interactive signing. The absence of reduction in monologue is used to argue that the dialogue compression reflects interactional entrainment rather than individual effort reduction.

Significance. If the claims are supported, the paper would provide valuable quantitative evidence on pragmatic adaptation in signed language and would highlight a concrete limitation of current sign-language models. The dataset itself is a notable contribution: dyadic ASL interaction captured with 3D motion capture, paired with a monologue from the same instructor, is rare and difficult to collect. The kinematic metrics are externally defined and the within-dialogue repeated-mention analysis is a sensible design. The authors are also honest about the case-study status, the small sample, and the hearing/non-native perspective in annotation. However, the central causal interpretation—distinguishing entrainment from effort reduction—depends on a monologue control whose statistical power and measurement compatibility are not established. The isolated-vocabulary baseline also conflates signer identity with context for the instructor.

major comments (2)

[§4.2.2 / §3.2] The load-bearing claim that repeated-mention compression is absent in monologue rests entirely on the null result r = 0.026, p = 0.895. The paper does not report how many repeated mentions or sign types entered this test, nor a confidence interval. A null correlation with unknown sample size is uninterpretable: it may reflect low statistical power, the 2D MediaPipe tracking (vs. Vicon 3D for the dialogue), or the hand-selected 'keyword overlap' excerpt, rather than a genuine absence of reduction. This issue is central because Contribution 2 and the Discussion/Conclusion use the monologue null to infer entrainment over effort reduction. Please report the token/type counts and a confidence interval for the monologue duration correlation, and preferably add a same-modality control or additional monologic data; otherwise, the causal claim should be explicitly softened.
[§4.1] The isolated vocabulary baseline was collected only from the student, yet it is used as the baseline for both participants ('the vocabulary signs produced by the student serve as the baseline reference for both participants'). Consequently, the headline duration reductions (instructor 24.6%, student 44.6%) conflate signer identity with discourse context. If the instructor naturally signs at a different rate or with a different spatial style than the student, these numbers do not measure a dialogue effect for the instructor. The authors should either collect isolated productions from the instructor as well, or explicitly restrict the quantitative cross-context comparison to the student and present the instructor's comparisons as exploratory.

minor comments (5)

[§3.4.1] The definition of Δcos is unclear: Δcos = cos(x1, xT) − cos(x1, x1) is described as 'cross-signer similarity,' but the formula as written appears to be within-signer (x1 and xT from the same signer). Please specify which signer each vector belongs to and define the intended comparison.
[§4.2.2] The statement 'r = 0.279, p < .001, indicating that signs become approximately 27.9% shorter' conflates the Spearman correlation coefficient with a percentage reduction. A correlation of 0.279 is not the mean percent reduction. Please report the actual mean percentage change separately from the correlation.
[Appendix A.2 / §4.2.2] Appendix A.2 states that for the student 'None of the correlations reached significance,' but §4.2.2 reports a significant student duration trend (r = 0.295, p < .001). If the appendix excludes duration, say so explicitly; otherwise the two statements are inconsistent.
[Table 2] The row labels 'Dinst.', 'D♢ inst.', 'Dstud.', 'Minst.', etc. are cramped and the meaning of the ♢ symbol is given only in the text. Please define all symbols in the caption and improve the formatting.
[Figure 2/3] The mean ± SE lines in Figures 2 and 3 are difficult to distinguish for individual signs; increasing the contrast or using a separate panel would improve readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; reported reductions are direct kinematic measurements compared against an external monologue control.

full rationale

The central quantities—sign duration, spatial extent, path length, velocity, vertical position—are physically defined metrics computed from Vicon/MediaPipe landmarks, not quantities defined in terms of the conclusions. The dialogue-vs-isolated comparison (§4.1) is a direct measurement against the student's citation-form productions. The repeated-mention analysis (§4.2) uses a first-mention baseline as a measurement convention, and the dialogue-vs-monologue contrast is an empirical control, not a fitted parameter. The monologue control is the weakest evidential link: it relies on a single Atomic Hands video of the same instructor, processed with MediaPipe while the dialogue used Vicon, and the null duration correlation (r = 0.026, p = 0.895; §4.2.2) is not reported with token counts or confidence intervals. That is a validity/power limitation and could threaten the entrainment interpretation, but it is not circularity: no equation is reused as a conclusion, and the null is not built into the kinematic definitions. The model evaluations use external pre-trained encoders (SignCLIP, I3D) as probes, and the sign-spotting task reports retrieval metrics against manually annotated ground truth; no fitted parameter is renamed as a prediction. The only author-overlapping sources (Atomic Hands video, Sem-Lex checkpoint, STEM-sign variation citation) are data/resources rather than self-justifying theorems, and the paper discloses that the same instructor appears in both dialogue and monologue. I find no load-bearing self-citation chain or definitional reduction; score 2 reflects the minor self-referential data source and the reliance on a weak but independent control, not actual circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no free physical parameters or invented entities: its central quantities (sign duration, path length, spatial extent, velocity) are measured directly. The load the paper does not pay for sits in the hand-chosen analysis choices (monologue excerpt selection, spotting thresholds, MediaPipe confidence) and in the assumptions above: the cross-signer vocabulary baseline, the cross-modality monologue comparison, the monologue's statistical power, annotation boundary accuracy, and token independence in pooled correlations.

free parameters (3)

Monologue excerpt selection = 8 sentences (Reproduction) + 12 sentences (Photosynthesis)
§3.2: monologue excerpts were selected 'to maximize keyword overlap with the dialogue'. This hand-chosen selection controls which sign tokens enter the null monologue analysis; with only two short excerpts the number of repeated mentions is small and the interpretation of the null result depends on it.
Sign spotting evaluation thresholds = window width = stride = 0.5s; IoU >= 0.3; k = 10, 50
§3.4.2: hand-chosen thresholds and ranks for the retrieval evaluation. Reported MRR and R@k values are conditional on these choices; no sensitivity analysis is given.
MediaPipe confidence thresholds = 0.5 (default)
§3.2: default detection and tracking confidence thresholds are used for the monologue and interpreter video conditions; landmark quality in these conditions depends on this choice.

assumptions (5)

ad hoc to paper Vocabulary baseline transferability: isolated vocabulary signs produced by the student serve as a valid articulation baseline for both participants, including the instructor.
§4.1: 'the vocabulary signs produced by the student serve as the baseline reference for both participants.' This equates signer identity with pragmatic context; if signer-specific articulation differs systematically, the instructor's 24.6% duration reduction is partly an artifact.
domain assumption Cross-modality kinematic comparability: kinematic features extracted from Vicon 3D motion capture (dialogue) are directly comparable to MediaPipe 2D landmarks (monologue, interpreter video) after manual joint mapping.
§3.2 and Appendix A.1: MediaPipe landmarks are mapped to motion capture joint names, but unit scales, 3D estimation accuracy, and tracking noise differ between modalities; percent-change statistics mitigate but do not eliminate this.
ad hoc to paper Monologue power sufficiency: the single Atomic Hands monologue video contains enough repeated mentions of the relevant STEM signs for the null duration result (r = 0.026, p = 0.895) to be interpretable as absence of reduction rather than low power.
§4.2.2 reports the null but not the number of monologue tokens or signs; the discriminator between entrainment and effort reduction depends on this null being informative.
domain assumption Annotation boundary accuracy: sign start/stop frames glossed by two hearing ASL-proficient researchers are accurate enough for duration comparisons.
§3.1 annotation protocol; the Limitations section itself acknowledges that 'judgments about translation correctness and sign start/stop frame are inherently subjective' and that hearing annotators may introduce bias.
standard math Token independence in pooled Spearman correlations: repeated-mention tokens from different signs can be pooled into a single rank correlation.
§4.2: correlations between mention index and percent reduction are computed across signs; if multiple tokens come from the same sign, the effective sample size is smaller than the token count suggests and reported p-values may be anti-conservative. Token counts per analysis are not reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of How Pragmatics Shape Articulation: A Computational Case Study in STEM ASL Discourse." pith.science (2026). https://pith.science/paper/KU7JNJD2

@misc{pith2026251023842,
  author       = {Pith},
  title        = {Pith review of: How Pragmatics Shape Articulation: A Computational Case Study in STEM ASL Discourse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KU7JNJD2}},
  note         = {Machine review of arXiv:2510.23842}
}

read the original abstract

Most state-of-the-art sign language models are trained on interpreter or isolated vocabulary data, which overlooks the variability that characterizes natural dialogue. However, human communication dynamically adapts to contexts and interlocutors through spatiotemporal changes and articulation style. This specifically manifests itself in educational settings, where novel vocabularies are used by teachers, and students. To address this gap, we collect a motion capture dataset of American Sign Language (ASL) STEM (Science, Technology, Engineering, and Mathematics) dialogue that enables quantitative comparison between dyadic interactive signing, solo signed lecture, and interpreted articles. Using continuous kinematic features, we disentangle dialogue-specific entrainment from individual effort reduction and show spatiotemporal changes across repeated mentions of STEM terms. On average, dialogue signs are 24.6%-44.6% shorter in duration than the isolated signs, and show significant reductions absent in monologue contexts. Finally, we evaluate sign embedding models on their ability to recognize STEM signs and approximate how entrained the participants become over time. Our study bridges linguistic analysis and computational modeling to understand how pragmatics shape sign articulation and its representation in sign language technologies.

Figures

Figures reproduced from arXiv: 2510.23842 by the authors.

**Figure 1.** We explore the effects of pragmatics in three different signed STEM contexts: 1) instructor-student dyadic conversations, 2) isolated vocabulary, 3) a signed lecture, and 4) interpreted Wikipedia articles. We computationally analyze how signers establish conceptual pacts across these contexts. To study this, we collect an American Sign Language (ASL) motion capture dataset that consists of two contexts: 1) an inst… view at source ↗

**Figure 2.** Sign duration patterns over time for instructor-student ASL dialogue. Stacked panels show start [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Path length differences between dialogue and vocabulary articulation for left and right hands as [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Sign productions plotted with respect to [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 3 linked inside Pith

[1]

Introduction Humancommunicationisinherentlyadaptive(Clark andBrennan,1991), causingcontext-sensitivephe- nomena such asentrainment, where speakers adjust their linguistic and articulatory behaviors to one another (Brennan and Clark, 1996), andeffort reduction, where individuals reduce articulatory effort through repetition (Zipf, 2016). While work in dial...

arXiv 1991
[2]

We find that dialogue signing exhibits reduced spatial and temporal articulation compared to isolated vocabulary productions, reflecting adaptive changes characteristic of interactive communication (§ 4.1)
[3]

We show statistically significant spatial and temporal reduction in STEM signs across re- peated mentions in dialogue, absent in mono- logue, indicating that such adaptation may re- flect entrainment rather than individual effort reduction (§ 4.2)
[4]

We demonstrate that existing sign language models fail to generalize to these articulatory differences (§ 4.4)
[5]

2011; Sigurd et al., 2004; Gibson et al., 2019)

Related Work Articulatory Reduction and Interactional Adap- tationA line of research argues that spoken lan- guagesareshapedbyafunctionalpressuretoward ease of articulation and communicative efficiency (Zipf, 2016; Kanwal et al., 2017; Piantadosi et al., 1Data can be made available to researchers upon proper agreements. 2011; Sigurd et al., 2004; Gibson e...

2016
[6]

Then, we describe our analysis of the data from the per- spective of motion capture kinematics (§3.3) and pretrained machine learning models (§3.4)

Methods Inthissection,wedescribeourmethodofcollecting and annotating the STEM dialogue data (§3.1) and finding comparable monologic data (§3.2). Then, we describe our analysis of the data from the per- spective of motion capture kinematics (§3.3) and pretrained machine learning models (§3.4). 3.1. Data Collection ParticipantsTwo fluent deaf signers partic...
[7]

Isolated STEM vocabulary articulation.One signer(thestudentdescribedabove)produced a set of 77 STEM signs drawn from introduc- tory biology content
[8]

ASL Signs – cell division, mi- tosis, meiosis

Biology dialogue.The two participants engaged in an 8.52-minute spontaneous instructor-student dialogue. The conversa- tion comprised four main question-answer ex- changes initiated by the instructor, focusing on keybiologytopicssuchascellstructureandge- netics, cell cycle, and photosynthesis. Across the dialogue, 17 STEM signs overlapped with the vocabul...

2025

Show all 20 references

[9]

Results We first compare dialogue articulations withVocab- ularybaselines to characterize how signing style differs across communicative contexts (§ 4.1). This analysis primarily uses left and right hand motion capture data, as the hands are the primary articula- tors that acc...

2010
[10]

Signs produced in dialogue were both spatially and temporally reduced relative to isolated sign articulations, with reductions inten- sifying across repeated mentions

Discussion This work provides the first quantitative evidence that pragmatic adaptation in sign language follows measurable articulatory principles comparable to spokendialogue(Zipf,2016;Kanwaletal.,2017;Pi- antadosi et al., 2011). Signs produced in dialogue were both spatiall...

2016
[11]

instructor

Conclusion This study presented an empirical analysis of how dialogic ASL signing differs from citation-style ar- ticulation through quantitative analyses of motion capturerecordinginSTEMdiscourse. Usingcontin- uous spatial, temporal, and vertical motion metrics, we found that...

2022
[18]

InInternational Conference on Machine Learning

Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning. Alex Shaw and Lisa Anthony. 2016. Analyzing the articulation features of children’s touchscreen gestures. InProceedings of the 18th ACM Inter- national Confer...

2016
[19]

InProceedings of the LREC2022 10th Workshop on the Rep- resentation and Processing of Sign Languages: Multilingual Sign Language Resources, pages 187–191

Capturing distalization. InProceedings of the LREC2022 10th Workshop on the Rep- resentation and Processing of Sign Languages: Multilingual Sign Language Resources, pages 187–191. Thad Starner, Sean Forbes, Matthew So, David Martin, Rohit Sridhar, Gururaj Deshpande, Sam Sepah,...

2023 arXiv
[20]

2022.SLAASh ID glossing principles, ASL Signbank and annotation conventions

Language Resource References Hochgesang, J. 2022.SLAASh ID glossing principles, ASL Signbank and annotation conventions. PID https://doi.org/10.6084/m9.figshare.12003732.v4. Jiang, Zifan and Sant, Gerard and Moryossef, Amit and Müller, Mathias and Sennrich, Rico and Ebling, Sa...

2022 arXiv
[174]

MayumiBono,TomohiroOkada,VictorSkobov,and Robert Adam

Springer. MayumiBono,TomohiroOkada,VictorSkobov,and Robert Adam. 2024. Data integration, annota- tion, andtranscriptionmethodsforsignlanguage dialogue with latency in videoconferencing. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Process- ing...

2024
[439]

Pengfei Lu and Matt Huenerfauth

Springer. Pengfei Lu and Matt Huenerfauth. 2012. CUNY AmericanSignLanguagemotion-capturecorpus: First release. InProceedings of the LREC2012 5th Workshop on the Representation and Pro- cessing of Sign Languages: Interactions be- tween Corpus and Lexicon, pages 109–116, Is- tan...

2012
[2014]

Matt Huenerfauth and Pengfei Lu

Do repeated references result in sign re- duction?SignLanguage&Linguistics, 17(1):56– 81. Matt Huenerfauth and Pengfei Lu. 2010. Elicit- ing spatial reference for a motion-capture corpus of american sign language discourse. Insign- lang@ LREC 2010, pages 121–124. European Lang...

2010
[2021]

spreadthe- sign

How2sign: a large-scale multimodal dataset for continuous american sign language. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735–2744. Simon Garrod and Anthony Anderson. 1987. Say- ing what you mean in dialogue: A study in con- ...

1987
[2022]

Claude E Mauk and Martha E Tyrone

Wmt-slt focusnews: Training data for the wmt shared task on sign language translation. Claude E Mauk and Martha E Tyrone. 2012. Loca- tion in asl: Insights from phonetic variation.Sign Language & Linguistics, 15(1):128–146. Claude Edward Mauk. 2003.Undershoot in two modalities...

2012
[2024]

InProceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of SignLanguageResources, pages 54–65, Torino, Italia

Systemic biases in sign language AI re- search: A deaf-led call to reevaluate research agendas. InProceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of SignLanguageResources, pages 54–65, Torino, Italia. ELRA an...

2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Introduction Humancommunicationisinherentlyadaptive(Clark andBrennan,1991), causingcontext-sensitivephe- nomena such asentrainment, where speakers adjust their linguistic and articulatory behaviors to one another (Brennan and Clark, 1996), andeffort reduction, where individuals reduce articulatory effort through repetition (Zipf, 2016). While work in dial...

arXiv 1991

[2] [2]

We find that dialogue signing exhibits reduced spatial and temporal articulation compared to isolated vocabulary productions, reflecting adaptive changes characteristic of interactive communication (§ 4.1)

[3] [3]

We show statistically significant spatial and temporal reduction in STEM signs across re- peated mentions in dialogue, absent in mono- logue, indicating that such adaptation may re- flect entrainment rather than individual effort reduction (§ 4.2)

[4] [4]

We demonstrate that existing sign language models fail to generalize to these articulatory differences (§ 4.4)

[5] [5]

2011; Sigurd et al., 2004; Gibson et al., 2019)

Related Work Articulatory Reduction and Interactional Adap- tationA line of research argues that spoken lan- guagesareshapedbyafunctionalpressuretoward ease of articulation and communicative efficiency (Zipf, 2016; Kanwal et al., 2017; Piantadosi et al., 1Data can be made available to researchers upon proper agreements. 2011; Sigurd et al., 2004; Gibson e...

2016

[6] [6]

Then, we describe our analysis of the data from the per- spective of motion capture kinematics (§3.3) and pretrained machine learning models (§3.4)

Methods Inthissection,wedescribeourmethodofcollecting and annotating the STEM dialogue data (§3.1) and finding comparable monologic data (§3.2). Then, we describe our analysis of the data from the per- spective of motion capture kinematics (§3.3) and pretrained machine learning models (§3.4). 3.1. Data Collection ParticipantsTwo fluent deaf signers partic...

[7] [7]

Isolated STEM vocabulary articulation.One signer(thestudentdescribedabove)produced a set of 77 STEM signs drawn from introduc- tory biology content

[8] [8]

ASL Signs – cell division, mi- tosis, meiosis

Biology dialogue.The two participants engaged in an 8.52-minute spontaneous instructor-student dialogue. The conversa- tion comprised four main question-answer ex- changes initiated by the instructor, focusing on keybiologytopicssuchascellstructureandge- netics, cell cycle, and photosynthesis. Across the dialogue, 17 STEM signs overlapped with the vocabul...

2025