Skip to content

Structured Claim-Level Discourse Representations for Dense Health Narratives

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: dense health narratives, discourse analysis, atomic claim decomposition, multidimensional pragmatic profiling, argument mining

TL;DR

To resolve multi-aspect entanglement in short-form health videos where speakers assert an average of 13.22 claims per minute, this paper introduces a structured discourse framework that maps raw conversational transcripts into atomic claim tuples \((c, a, s, \vec{\phi})\) across domain aspects, stances, and a 6-axis pragmatic typology, revealing critical trade-offs between local semantic anchoring and global context degradation.

Background & Motivation

Short-form video platforms such as YouTube Shorts and TikTok have become dominant sources for public health communication, yet user-generated health narratives in these environments remain informal, rhetorically volatile, and computationally elusive. Existing computational approaches primarily rely on coarse document- or segment-level labels, including broad topic classification, overall sentiment analysis, or stance detection (Augenstein et al., 2016). While suitable for well-edited written essays or structured public debates, these coarse formulations fail when confronted with spontaneous social media narratives, where speakers rapidly interlace physiological mechanisms, personal recovery anecdotes, drug dosage protocols, and subjective recommendations within very brief conversational spans. Empirical quantification across real-world datasets reveals that these videos contain an average of 13.22 atomic propositions per minute, manifesting a severe phenomenon of multi-aspect entanglement.

The central tension arises because a single continuous discourse segment frequently bundles diametrically opposing communicative stances and heterogenous pragmatic framing. In testosterone replacement therapy (TRT) narratives, for example, a creator may tout muscular hypertrophy gains (positive stance) while in the very next breath warning of severe cardiovascular infarction and permanent endocrine shutdown (negative stance), shifting fluidly from prescriptive lifestyle policy to personal anecdotal narrative and cited statistical literature. Flat document-level stance or topic assignments completely collapse these vital distinctions, obscuring the boundary between therapeutic benefits and adverse toxicities. Even fine-grained Aspect-Based Sentiment Analysis (ABSA) falls short because it only captures positive or negative sentiment towards a target entity, offering zero representational capacity for underlying causal mechanics, epistemic certainty, empirical authority, or persuasive intent.

To overcome the representational collapse of flat NLP paradigms in dense health discourse, this paper formulates fine-grained claim-level discourse analysis as a structured decomposition problem. The authors break compound narrative transcripts down into minimal self-contained propositional units and jointly model their domain-specific thematic aspect, directional stance, and an orthogonal 6-axis pragmatic discourse profile. Core idea: decompose dense health discourse into structured tuples \((c, a, s, \vec{\phi})\) uniting atomic claims with inductive thematic aspects, stances, and domain-invariant multidimensional pragmatic vectors, capturing granular rhetorical trajectories that coarse document- and sentiment-level models obscure.

Method

Overall Architecture

The proposed framework operates as a structured discourse parsing and profiling pipeline designed to process unstructured spoken conversational video narratives. Raw audio streams are first retrieved and converted to text via automated speech recognition (ASR); the resulting text is passed through an orthographic denoising and boundary stabilization stage. The stabilized transcript is then segmented into self-contained atomic claims with explicit anaphora grounding. Finally, context-guided inference models predict the domain-specific thematic aspect, core stance polarity, and a 6-axis domain-invariant pragmatic typology vector for each atomic claim.

The complete workflow is illustrated in the vertical flowchart below:

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Raw Health Video Audio<br/>Retrieved via YouTube Data API"] --> B["ASR Transcription & Orthographic Denoising<br/>Whisper + Gemini 2.5 Flash Stabilization"]
    B --> C["Atomic Claim Decomposition<br/>Clause Splitting & Anaphora Grounding"]
    C --> D["Inductive Thematic Aspect Modeling<br/>Domain-Adaptive Taxonomy Alignment"]
    D --> E["6-Axis Pragmatic Typology Profiling<br/>Logic / Verifiability / Evidence / Certainty / Time / Focus"]
    E --> F["Structured Discourse Tuples<br/>T_V = {(c, a, s, phi)_i}"]

Key Designs

1. ASR Transcription Denoising and Atomic Claim Decomposition: Overcoming Transcript Fragmentation and Propositional Entanglement

Spontaneous spoken transcripts generated by speech-to-text models (Whisper) frequently suffer from severe orthographic degradation, fragmented sentence boundaries, and corrupted medical terminology (e.g., transcribing "Ozempic" as "a Zempic" and "autophagy" as "a topogy"). To prevent cascading errors in downstream semantic parsing, the authors implement an Orthographic Denoising and Stabilization layer using Gemini 2.5 Flash under conservative constraints that strictly prohibit summarization, deletions, or narrative insertions. Following text stabilization, the pipeline performs Atomic Claim Decomposition: complex conversational segments are split into minimal propositional units \(c_i\) that each express a single coherent assertion. Unresolved pronouns and deictic markers are grounded using anaphora resolution principles, ensuring that each atomic proposition is fully self-contained and interpretable outside of its local conversational clause.

2. Inductive Thematic Aspect Modeling: Capturing Dynamic Semantic Spaces Across Health Domains

Health domains exhibit vastly divergent clinical concerns and vernacular semantics, rendering rigid, predefined universal taxonomies ineffective. To establish ecologically valid semantic categories, the framework employs an inductive aspect construction methodology. Using LLM-assisted claim clustering over pilot narrative transcripts followed by expert human adjudication, the authors induce domain-tailored aspect inventories \(\mathcal{A}\) for four major health topics: Ozempic/GLP-1 (13 aspects, spanning appetite suppression, gastrointestinal adverse effects, and financial access), TRT (14 aspects, including muscle aesthetics, sexual health, HPTA shutdown, and clinical eligibility), Collagen (12 aspects, covering skin anti-aging, joint mobility, and formulation), and Intermittent Fasting (19 aspects, including autophagy, metabolic rate, and feeding protocols). This allows every extracted claim \(c_i\) to be categorized into an interpretable, domain-specific semantic target.

3. 6-Axis Pragmatic Typology Profiling: Disentangling Domain-Invariant Epistemic and Communicative Structures

While thematic aspects identify what a claim is about, they provide no information regarding how the claim is framed, how it is defended, or what communicative purpose it serves. The paper addresses this by formalizing an orthogonal six-dimensional pragmatic vector \(\vec{\phi} = (\phi_1, \dots, \phi_6) \in \mathcal{D}_1 \times \dots \times \mathcal{D}_6\): - Logical Relationship: The underlying structural reasoning mechanism connecting concepts, classified as Causal, Correlational, Comparative, Conditional, or None_Logical; - Verifiability & Intent: The epistemological nature of the statement, spanning Factual (Clinical/Biological), Factual (Logistical/Contextual), Value (Subjective), Policy (Prescriptive), and Speculative/Rumor; - Evidence Basis: The authoritative ground invoked to substantiate the claim, categorized into Scientific (Statistical), Scientific (Expert/Authority), Scientific (General/Vague), External Reference (Media/Cultural), Anecdotal/Personal, Visual/Observable, General Knowledge/Common Sense, or Commercial; - Certainty: The speaker's epistemic commitment and modality, spanning Absolute/Imperative, Hedged (Probabilistic), and Attributed/Distanced; - Temporality: The temporal orientation or universality of the claim, classified into Current/Past, Counterfactual, Future/Predictive, Sequential, and Timeless; - Claim Focus: The primary communicative function in the narrative arc, including Explanatory/Argumentative, Descriptive/Definitional, Narrative/Sequential, Evaluative, and Promotional/Commercial. This 6-axis profiling enables the formal tracking of intra-speaker rhetorical trajectories across continuous video timelines.

4. Task-Decomposed Contextual Tuning: Mitigating Long-Range Contextual Dilution in High-Dimensional Inference

To resolve cognitive interference in large language models when jointly predicting heterogeneous discourse variables, the authors examine context-window dynamics across tasks. Their analysis demonstrates that while thematic aspect and stance detection thrive in local batch settings (Batch Window Size = 8) with category definitions and keyword anchors, high-dimensional pragmatic classification degrades severely when exposed to broad contextual windows. Providing full transcripts or compressed summaries introduces conversational noise that dilutes the model's focus on granular sentence mechanics. Consequently, the optimal inference architecture pairs a localized context window (Local \(\pm 2\) claims) with explicit heuristic discourse constraints derived from human annotation reconciliation sessions. These linguistic rules anchor structural transitions and connective cues, allowing models to bypass contextual over-analysis.

Benchmark & Evaluation Protocol

To benchmark automated discourse understanding, the authors constructed a gold-standard dataset across 60 curated YouTube videos (15 per domain across Ozempic, TRT, Collagen, and Fasting), encompassing 652 complex statements and 1,191 decomposed atomic claims across 90 minutes of content. Two computational discourse experts spent approximately 400 person-hours executing a four-stage annotation and consensus adjudication process. Reliability metrics demonstrate exceptional rigor: statement extraction attained a pairwise \(F_1\) of 0.93; atomic decomposition achieved 92% raw agreement; aspect assignment and stance classification achieved Cohen's \(\kappa\) of 0.83 and 0.80, respectively; and the 6-axis typology achieved an average agreement of 88% (\(\kappa = 0.76\)). This dataset serves as the standard evaluation benchmark for both proprietary and open-weight language models.

Key Experimental Results

Main Results

The main benchmark evaluated multiple proprietary and open-weight language models under identical pipeline prompting across all four health domains on Aspect Labeling, Stance Detection, and 6-Axis Typology classification.

Model Domain Aspect Acc. (%) / \(\kappa\) Stance Acc. (%) / \(\kappa\) 6-Axis Typology Mean Acc. (%) / \(\kappa\)
Gemini 2.5 Flash Ozempic 79.23 / 0.77 91.55 / 0.87 86.15 / 0.76
Gemini 2.5 Flash TRT 89.12 / 0.88 92.15 / 0.88 88.62 / 0.77
Gemini 2.5 Flash Collagen 86.44 / 0.85 85.59 / 0.77 83.97 / 0.72
Gemini 2.5 Flash Fasting 77.64 / 0.75 83.38 / 0.73 82.23 / 0.59
Qwen 3.5 35B Ozempic 77.54 / 0.75 85.26 / 0.78 81.58 / 0.70
Qwen 3.5 35B TRT 79.00 / 0.77 83.30 / 0.74 82.05 / 0.62
Qwen 3.5 35B Collagen 81.36 / 0.79 78.39 / 0.65 81.43 / 0.68
Qwen 3.5 35B Fasting 78.25 / 0.76 86.71 / 0.79 80.72 / 0.58
Llama 3.3 Ozempic 76.49 / 0.74 83.86 / 0.75 75.56 / 0.61
Llama 3.3 TRT 64.77 / 0.61 82.56 / 0.74 73.90 / 0.49
Llama 3.3 Collagen 74.15 / 0.71 75.85 / 0.60 71.96 / 0.52
Llama 3.3 Fasting 67.37 / 0.64 84.59 / 0.75 69.79 / 0.41
Gpt-oss 20B Ozempic 78.60 / 0.76 81.40 / 0.72 73.04 / 0.59
Gpt-oss 20B TRT 71.17 / 0.68 82.56 / 0.73 75.33 / 0.53
Gpt-oss 20B Collagen 71.61 / 0.68 79.66 / 0.66 75.85 / 0.59
Gpt-oss 20B Fasting 70.69 / 0.67 87.31 / 0.80 76.08 / 0.52

Ablation Study

The ablation study systematically evaluated discourse context settings alongside heuristic rules and Chain-of-Thought (CoT) prompting using Gemini 2.5 Flash on the 6-Axis Typology task.

Context & Prompting Configuration Logical Rel. (I) Verifiability (II) Evidence (III) Certainty (IV) Temporality (V) Claim Focus (VI) Mean Acc. (%) / \(\kappa\)
Atomic (Isolated Claim) 55.91% 82.80% 79.57% 93.91% 71.33% 52.69% 72.70% / 0.57
Batch (Chronological 8-Claim Window) 57.71% 81.72% 62.37% 92.47% 63.80% 53.05% 68.52% / 0.53
Local \(\pm 2\) (Immediate 2 Neighbors) 55.56% 84.59% 77.78% 91.40% 62.72% 53.76% 70.97% / 0.54
Narrative (High-Level Arc Summary) 54.84% 84.95% 58.78% 90.32% 66.67% 51.61% 67.86% / 0.52
Transcript (Full Video Transcript) 57.71% 81.00% 44.44% 86.38% 60.93% 46.24% 62.78% / 0.43
Summary (Compressed Video Summary) 57.35% 82.08% 53.76% 92.11% 68.82% 41.22% 65.89% / 0.49
Local \(\pm 2\) + Heuristics 84.23% 86.38% 84.23% 90.32% 91.40% 82.08% 86.44% / 0.76
Local \(\pm 2\) + Heuristics + CoT 80.29% 82.08% 61.65% 91.04% 74.19% 49.46% 73.12% / 0.59

Key Findings

  • Contextual Noise and Semantic Dilution: Providing full transcripts as context degrades pragmatic discourse profiling substantially (dropping mean accuracy from 72.70% in the isolated Atomic setting to 62.78%). Most critically, Evidence Basis accuracy collapses by 35.13 percentage points (from 79.57% to 44.44%), demonstrating that broader conversational spans distract LLMs from identifying the exact evidentiary authority grounding an individual assertion.
  • Pathological Failure of CoT in Multidimensional Pragmatics: While lightweight CoT improves stance detection, adding CoT to pragmatic classification severely degrades performance (mean accuracy plummets from 86.44% to 73.12%). Qualitative error analyses reveal three failure modes: joint prediction interference across orthogonal axes, semantic over-analysis ignoring schema boundaries (e.g., explaining a superlative correctly but labeling it Non_Logical instead of Comparative), and autoregressive rationalization traps where modal words ("can") force the model to justify incorrect probabilistic hedges.
  • Open-Weight Capability Gap on Pragmatic Reasoning: While open-weight models such as Qwen 3.5 35B closely match proprietary performance on aspect and stance classification (within 1–3% of Gemini 2.5 Flash), their performance deteriorates sharply on the 6-axis pragmatic typology task, where smaller models (Llama 3.3, Gpt-oss 20B) exhibit Cohen's \(\kappa\) scores dropping below 0.50.

Highlights & Insights

  • Quantifying Conversational Entanglement: The paper empirically benchmarks the extreme density of social media health discourse (13.22 claims/min), establishing that conventional document-level classification fails to capture modern conversational communication.
  • Decoupled Thematic vs. Pragmatic Typology: Introducing an orthogonal 6-axis typology demonstrates that claims sharing identical aspect and stance can diverge across all six pragmatic axes, providing critical granularity for automated misinformation detection.
  • The Local Context Principle in Pragmatics: The findings debunk the assumption that wider conversational context uniformly aids discourse parsing, proving that a localized window (Local \(\pm 2\)) paired with explicit structural heuristics achieves superior pragmatic grounding.

Limitations & Future Work

  • Unimodal Spoken Modality: The benchmark is constructed purely from speech-to-text transcripts, excluding visual demonstrations, paralinguistic tone, and on-screen graphical overlays that often qualify or contradict spoken medical claims.
  • Heuristic Transferability Constraints: The top-performing configuration relies on hand-crafted heuristic rules derived from adjudication sessions, which may require adaptation when applied to out-of-domain conversational datasets.
  • Single-Speaker Monologue Scope: The current dataset analyzes single-creator video narratives, omitting multi-speaker dialogue phenomena like conversational turn-taking, inter-speaker interruptions, and collaborative consensus building.
  • vs Aspect-Based Sentiment Analysis (ABSA): Prior ABSA systems (Cai et al., 2021; Dragut et al., 2010) map sentiments to specific entity targets but lack mechanisms to classify causal structures, empirical evidence types, and temporal assertions. The proposed tuple \((c, a, s, \vec{\phi})\) unifies entity-level targeting with multi-axis pragmatic profiling.
  • vs Argument Mining & Evidentiality: Traditional argument mining benchmarks (Habernal & Gurevych, 2017; Mochales & Moens, 2011) focus on formal debates or student essays with low claim density and clean syntax. This work successfully adapts argument and evidentiality schemas to high-density, unscripted social media video discourse.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates the novel task of structured claim-level discourse modeling for dense narratives with a rigorous 6-axis pragmatic schema.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Implements extensive context ablations across 4 health domains, 60 videos, 1,191 gold claims, and multi-temperature stability trials.
  • Writing Quality: ⭐⭐⭐⭐⭐ Highly coherent, conceptually rigorous, and supported by rich qualitative error breakdowns.
  • Value: ⭐⭐⭐⭐⭐ Provides vital infrastructure for social media fact-checking, medical misinformation tracking, and nuanced health NLP.