Semantic Action Graph: A Shared Representation for Agent Grounding and Human Interpretation of Sports Highlights
Tica Lin, Deepak Chandran, Gauri Jagatap, Chen Chen, Andrea Fanelli, David Gunawan, Josh Kimball
cs.HC · cs.AI
Quick verdict
Proposes a graph schema grounding both LLM highlight agents and viewer interfaces; promising probe, but single-match, no-baseline study limits claims.
Summary
The paper introduces the semantic action graph, a lightweight soccer-match schema of five node types (performer, action, recipient, moment, state) connected by five edge types (perform, affect, happen, produce, then). It is designed to serve two consumers simultaneously: an agentic highlight-generation pipeline and a viewer-facing graph interface, with three claimed properties — connected event sequences, a shared closed vocabulary, and frame-addressable moments. The authors instantiate it in SportSAGE, a design probe whose four-module pipeline (a deterministic Game Analyzer plus three Claude 3.5 Sonnet agents for narrative, clip assembly, and ranking) operates solely on official DFL league event data; a 96-minute Pokal Cup match yields 1,330 action nodes across 13 action types. A formative interview study with six fans identified four gaps (content, narrative, personalization, discovery), distilled into three design goals. Feedback from 12 soccer fans was positive (overall quality M=5.25/7 for both default and customized sets; preference for the customized narrative M=5.92; interface usefulness M=5.96) with reliable task completion (12/12 filtering, 11/12 locating). Main limitations, largely acknowledged by the authors: evaluation on a single match with one model, no unstructured-representation baseline, small samples with only means reported, and reliance on professional event feeds rather than video.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A small closed schema (five node types, five edge types) can simultaneously ground an LLM agentic highlight pipeline and drive a viewer-facing inspection interface.
- Connected event sequences, a shared closed vocabulary, and frame-addressable moments let one representation address content, narrative, personalization, and discovery gaps identified with fans.
- Deterministic statistics plus frame-index-derived clip boundaries keep agent output factual rather than hallucinated.
- Twelve soccer fans rated SportSAGE highlights and the graph interface positively (quality M=5.25/7, customized-narrative preference M=5.92, usefulness M=5.96) and completed verification tasks reliably (12/12, 11/12, 11/12).
Strengths
Coherent dual-consumer design traceable to user needs
The three schema properties (Sec. 4.2) map one-to-one onto design goals G1-G3 from the six-fan formative study (Sec. 3), which surfaced concrete gaps with quotes (e.g., F3 wanting build-up, F1 on missing red-card context) — the schema is motivated, not asserted.
Anti-hallucination engineering in the pipeline
Sec. 5.1: the Game Analyzer is deterministic ('removing arithmetic from the language model', citing Toolformer), and clip boundaries come from Moment frame indices rather than generated timecodes; agents never read video, so the graph is the sole factual interface (Fig. 3).
Explicit capability boundary via closed vocabulary
Task-based evaluation, not just Likert means
Candid limitation statement
Weaknesses
No baseline condition for the central causal claim
Personalization shows no quality gain
Single match, single model, single proprietary feed
Minimal statistical reporting
No error analysis of graph construction
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is an IEEE VIS 2026 workshop paper (GenAI, Agents, and the Future of VIS) from Dolby Laboratories. It introduces the **semantic action graph**, a lightweight domain schema representing a soccer match via five node types — performer, action, recipient, moment, state (Sec. 4.1) — connected by five edge types: perform, affect, happen (roles within one event) and produce, then (cross-event structure). The central claim is that one small schema can serve **two consumers at once**: an LLM agentic highlight-generation pipeline and a viewer-facing interface. Three properties are claimed (Sec. 4.2) — connected event sequences, a shared closed vocabulary, and frame-addressable moments — each mapped onto a design goal (G1–G3) derived from a formative study with six fans (Sec. 3) that surfaced four gaps (content, narrative, personalization, discovery). Contributions: the schema, the SportSAGE design probe, and user feedback (n=12).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 is the semantic action graph?
It is a lightweight schema for representing sports match events as a graph with five node types (performer, action, recipient, moment, state) and five edge types (perform, affect, happen, produce, then). It is designed so that the same structure can drive both an AI highlight-generation pipeline and a viewer-facing interface for searching and verifying highlights.
What is SportSAGE?
SportSAGE is a design-probe system from Dolby Laboratories that instantiates the semantic action graph for soccer. It pairs a four-module pipeline (deterministic match statistics plus three Claude 3.5 Sonnet agents for narrative, clip assembly, and ranking) with a Cytoscape.js graph interface, and generates personalized narrated highlights from official league event data.
How well did SportSAGE perform in its user study?
Twelve soccer fans rated overall highlight quality 5.25 out of 7 for both default and personalized sets, preferred the personalized narrative (5.92/7), rated the interface useful (5.96/7), and completed graph tasks reliably: 12/12 filtered shot events and 11/12 located a specific player action.
What are the main limitations of this paper?
Evaluation covers a single soccer match using one LLM and official DFL event data rather than video-derived analysis. There is no unstructured-representation baseline, so benefits cannot yet be attributed to the shared representation itself. Samples were small (6 formative interviews, 12 evaluation participants) and only mean ratings were reported, without variance or statistical tests.
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.
- Figures 1-3 are available only as captions in the extracted text; visual encodings, screenshots, and the graph rendering could not be assessed.
- Evaluation-session protocol details (task instructions, ordering, counterbalancing, apparatus) are not present in the text.
- No appendix content on prompts, narrative templates, or the complete 13-action catalog definition beyond enumerated names.
- Claims about feed granularity (2-5 s) and node counts could not be verified against underlying data.
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.