Evidano is an AI-powered qualitative data analysis platform that ingests, codes, and visualizes social media and transcript data. A June 30, 2026 PLOS ONE paper introduces ST-TAN, a RoBERTa-based model that fuses sentiment, topic, and temporal attention to automate social media narrative analysis (PLOS ONE). On the Twitter Event Narrative Dataset (TEND, n=120, 000) ST-TAN reaches a narrative F1 of 0.87 and narrative integrity of 85.3%, outperforming baselines by about 8–11%. If qualitative teams analyze social feeds, the paper shows how joint modeling reduces failure modes such as fragmented posts, emotion-driven topic shifts, and weak temporal ordering. Below: what the model does, practical implications, and a compact six-step workflow to reproduce the approach using AI-enabled tools like Evidano.
Key Takeaways
ST-TAN jointly models sentiment, topic, and temporal attention to automate narrative recognition on social media, and it shows measurable gains on a held-out-event benchmark. On the Twitter Event Narrative Dataset (TEND, n=120, 000) ST-TAN reports narrative F1 0.87 and narrative integrity 85.3%, improving baselines by roughly 8–11%. Qualitative teams can reproduce ST-TAN-style insights with a six-step pipeline: ingest, clean and segment, auto-code, validate, visualize, and share.
- On TEND (120, 000 tweets) ST-TAN achieves narrative F1 0.87 and narrative integrity 85.3%, outperforming baselines by approximately 8–11%.
- ST-TAN’s core innovations are joint sentiment-topic training, attention-weighted topic modeling, and a temporal-consistency loss that preserves narrative ordering.
- A practical six-step workflow (ingest, clean/segment, auto-code, validate, visualize, share) can reproduce ST-TAN-style outputs using tools like Evidano.
Findings Snapshot
| Date | Paper / Model | Dataset | Samples | Narrative F1 | Narrative integrity | Source |
|---|---|---|---|---|---|---|
| 30 Jun 2026 | ST-TAN (Pei et al.) | TEND (Twitter Event Narrative Dataset) | 120, 000 | 0.87 | 85.3% | PLOS ONE |
What Happened: ST-TAN in plain English
This section explains ST-TAN in plain English: ST-TAN is a Sentiment-Topic fused Temporal Attention Network built on a RoBERTa-base encoder that jointly models sentiment, topic, and temporal structure to detect narrative units in social streams. The authors build ST-TAN from a RoBERTa-base encoder plus three fused modules: an emotion-aware head, a topic attention head weighted by sentiment, and an attention-based temporal modeling head; key innovations include joint sentiment-topic training, a temporal consistency loss, and preprocessing that segments narrative units by 24-hour temporal windows plus semantic similarity.
- Dataset and design: the authors use TEND, a dataset of 120, 000 tweets across 10 public events, and they use held-out events for the train/val/test split to evaluate cross-event generalization.
- Performance highlights: ST-TAN reports narrative F1 0.87 (about +8.75% versus the best baseline), narrative integrity 85.3% (+10.2%), sentiment accuracy around 0.89, and topic accuracy around 0.90.
- Robustness: ST-TAN degrades less under simulated noise, with high-noise narrative F1 about 0.76, attributed to targeted preprocessing and sentiment-weighted losses.
How ST-TAN works (nuts & bolts)
This section summarizes the core pipeline and why each part matters: the pipeline first cleans and segments text into narrative units using a 24-hour temporal threshold plus semantic similarity, then encodes segments with RoBERTa, applies an emotion-aware head, runs sentiment-weighted topic attention, and applies temporal attention with a temporal-consistency loss during LoRA fine-tuning.
- Cleaning and segmentation: narrative-unit segmentation uses 24-hour windows and semantic similarity threshold (≥0.5).
- RoBERTa encoder: the model uses a 12-layer, 768-d RoBERTa base encoder.
- Joint modeling benefits: modeling sentiment and topic together helps detect topic transitions because sentiment shifts often signal changes in topic; removing temporal attention drops temporal-position F1 by about 14.9%.
- Annotation quality: Cohen’s kappa greater than 0.79 across tasks supports stable supervised signals for multi-task training.
Implications for qualitative researchers
For UX & product teams
This subsection explains why joint modeling matters for UX and product teams: joint sentiment-topic-time analysis surfaces not only what customers feel but when complaints cluster into distinct stages such as discovery, amplification, and resolution. Use joint modeling to identify role shifts, who is framing the issue, and the key events that drive churn or support requests.
For policy, comms, & crisis teams
This subsection explains the benefits for policy, communications, and crisis teams: temporal coherence enables reconstruction of narrative arcs, showing emergence, peak, and decline, which supports timely interventions and after-action reviews. The authors report modest cross-event generalization drops of about 5–12%, so trained models can transfer to unseen crises with caution and human oversight.
For qualitative methodologists
This subsection explains methodological implications: combining coding dimensions in model architecture reduces downstream reconciliation work by producing fewer conflicting labels across sentiment, topic, and time. The authors note limitations such as very short texts, pronoun-based coreference, and multi-event posts as persistent failure modes that require human validation.
Do more, faster with Evidano
Problem: messy, multilingual social feeds
This subsection explains the solution for messy social feeds: Evidano provides website and social scraping plus preprocessing including emoji handling and spell correction to match ST-TAN cleaning steps before modeling.
Problem: needing joint sentiment-topic-temporal coding
This subsection explains the solution for joint coding: Evidano can import transcripts and tweets and run thematic, sentiment, and cross-segment analyses together; Evidano supports hierarchical codes to subcodes and co-occurrence networks to replicate the joint insights ST-TAN surfaces.
Problem: scale and reproducibility
This subsection explains the solution for scale and reproducibility: Evidano supports LoRA-style efficient fine-tuning and repeatable pipelines where teams can save preprocessing steps, codebooks, and visual reports. Data is encrypted and never used to train third-party models.
Problem: stakeholder reporting
This subsection explains the solution for stakeholder reporting: Evidano provides clickable quotes, word clouds, temporal co-occurrence visualizations, and exportable reports that show narrative arcs and integrity metrics for decision makers.
This Week’s 6-Step Workflow to reproduce ST-TAN-style insights in Evidano
This section lists a six-step workflow to reproduce ST-TAN-style insights in Evidano.
1) Ingest: scrape the public social stream or upload a CSV of posts and timestamps using Evidano.
2) Clean and segment: apply Evidano preprocessing, including emoji mapping and spell correction, then segment by 24-hour windows and semantic clustering.
3) Auto-code: run combined sentiment and thematic analysis with Evidano AI models to get initial labels and confidence scores.
4) Validate: sample low-confidence or short-text cases, correct labels with human annotators, and push corrections back to retrain models.
5) Visualize: generate co-occurrence networks, timeline views, and narrative integrity metrics to trace arcs and high-impact nodes.
6) Share and act: export a stakeholder brief and set alerts for negative sentiment spikes or new topic emergence.
Wrapping up & next step
This section summarizes the main findings and the next step: ST-TAN demonstrates that integrating sentiment, topic, and temporal modeling materially improves automated narrative recognition on social media, as reported in PLOS ONE on 30 Jun 2026. For qualitative teams, the practical takeaway is to combine dimensions early in the pipeline and to bake temporal consistency into sequence-level analyses.
Ready to test a pipeline? Ingest your social feed or transcripts into Evidano and run a joint sentiment, topic, and temporal analysis in hours instead of weeks. Try Evidano for free.
FAQ: social media narrative analysis
What is ST-TAN and what problem does it solve?
ST-TAN is a Sentiment-Topic fused Temporal Attention Network that jointly models sentiment, topic, and temporal attention to automate narrative recognition in social media streams. The model combines a RoBERTa-base encoder with an emotion-aware head, sentiment-weighted topic attention, and temporal attention with a temporal-consistency loss to detect coherent narrative units.
How well does ST-TAN perform on standard benchmarks?
ST-TAN reports narrative F1 0.87 and narrative integrity 85.3% on the Twitter Event Narrative Dataset (TEND, n=120, 000). The authors also report sentiment accuracy around 0.89 and topic accuracy near 0.90, with robustness under noise (high-noise F1 near 0.76).
Can teams reproduce ST-TAN-style outputs with off-the-shelf tools?
Yes, teams can reproduce ST-TAN-style outputs by following the six-step pipeline: ingest, clean and segment, auto-code, validate, visualize, and share. Tools that provide scraping, preprocessing, hierarchical coding, repeatable pipelines, and visualization, such as Evidano, can implement this workflow.
What are the main limitations to watch for when using ST-TAN-style models?
The primary failure modes are very short texts, pronoun-based coreference, and multi-event posts, which can break narrative segmentation and labeling. The authors recommend human validation for these cases and caution when transferring models to unseen events, since cross-event generalization shows modest drops (about 5–12%).