Faster Narrative Insight: qualitative analysis of social media

ST-TAN demonstrates that fusing sentiment, topic, and temporal signals materially improves automated narrative parsing on social media, achieving narrative F1 of 0.87 on a 120, 000-tweet dataset with narrative integrity at 85.3%.

Evidano is an AI-powered qualitative data analysis platform that supports time-aware thematic workflows and the same preprocessing and joint-analysis steps described in the study.

• ST-TAN: narrative F1 = 0.87 and narrative integrity = 85.3% on the TEND dataset (n=120, 000).
• Model design: RoBERTa encoder with sentiment-weighted topic attention and temporal attention, plus gating fusion and joint losses.
• Practical workflow: apply 24-hour window segmentation, semantic similarity threshold ≈0.5, and a 1–2k annotated seed for LoRA-style fine-tuning.
• Ethics and security: study authors applied anonymization and discuss bias, and the study frames outputs as research tools that require human oversight.

Pei, Lyu & Wang (Published June 30, 2026) propose ST-TAN, a joint sentiment-topic-temporal Transformer that outperforms prior baselines on narrative recognition, sentiment and topic classification.

Read the original study at PLOS One.

• Key gains: narrative F1 = 0.87; narrative integrity = 85.3%; sentiment accuracy = 89%; topic accuracy = 90% (TEND dataset, n=120, 000).
• Core idea: sentiment-weighted topic attention, temporal attention, and a RoBERTa encoder, with practical preprocessing for fragmented social texts.
• Ethics note: authors applied anonymization, discuss bias and responsible use, outputs are research tools requiring human oversight.

ST-TAN treats sentiment, topic and time as interdependent signals rather than separate outputs.

ST-TAN uses RoBERTa (12-layer, 768 hidden) for encoding, then applies a sentiment-aware topic attention and a temporal attention module with a gating fusion.

Preprocessing segments posts into narrative units using a 24-hour temporal window and semantic similarity ≥0.5, standardizes emojis and corrects noisy text, and annotators labeled role/event/sentiment/topic to train the joint model.

• Joint losses: sentiment-weighted cross-entropy, topic cross-entropy, focal loss for narrative elements, plus a temporal-consistency loss.
• Fine-tuning: LoRA used to reduce parameter updates (≈90% parameter reduction during adaptation).
• Ablation shows temporal attention matters most: removing it drops narrative F1 by ≈9.2% and temporal position F1 by ≈14.9%.

Follow this run-book to reproduce ST-TAN style outputs in an AI-enabled qualitative workflow:

• 1) Collect: Pull platform data with timestamps and reply/retweet relationships.
• 2) Clean: Remove urls and handles, standardize emojis, correct spelling; preserve sentiment tokens via custom dictionary.
• 3) Segment: Group into narrative units using a 24-hour rolling window and a semantic similarity threshold (≈0.5).
• 4) Seed labels: Hand-annotate a balanced seed (roles, events, sentiment, topics) of ≈1–2k samples for domain adaptation.
• 5) Fine-tune: Use lightweight adaptation (LoRA-style) to specialize your encoder on domain language.
• 6) Run joint analysis: Generate thematic, sentiment, temporal, and co-occurrence outputs; validate top errors manually.
• 7) Deliver: Export visual reports, clickable quotes, and an executive timeline; iterate with stakeholders.

ST-TAN demonstrates that fusing sentiment, topic and temporal signals materially improves automated narrative parsing on social media (n=120, 000; narrative F1 0.87; integrity 85.3%).

For teams doing qualitative analysis of social media narratives, adopt time-aware joint models, keep tight preprocessing, and pair AI outputs with human review.

Ready to operationalize this? Upload a pilot corpus or connect your social feed to run a time-aware thematic analysis in minutes, or Try Evidano for free.

If you want a reproducible pilot, start with a 1–2k annotated seed and use Evidano’s AI-assisted coding and visual exports to produce timelines and stakeholder-ready reports.