WhatYouSay

1) Overview

• WhatYouSay uses a two-level analysis pipeline for exported text conversations.
• Level A produces deterministic, reproducible metrics from raw text.
• Level B generates a narrative interpretation via a large language model, grounded in Level A metrics.
• The levels are separated by design: Level A is independently verifiable; Level B is interpretive and should be read as pattern description, not diagnosis.

2) Level A: Lexicon-Based Analysis

Input: raw .txt file parsed into individual messages attributed to speakers.

Metrics produced:

• Message counts, character counts, and estimated word counts per speaker.
• Speaker participation distribution (percentage of corpus).
• Average message length (words per message).
• Question rate (percentage of messages containing question marks).
• Emoji usage rate.
• Overall tone valence (composite score).
• Emotion distribution via NRC Word-Emotion Association Lexicon (Mohammad and Turney, 2013).
• Moral/value framing via LibertyMFD (Araque et al., 2022) and MoralStrength (Araque et al., 2019).

Reliability thresholds:

• Individual analysis is most reliable with at least 250 messages from the selected user.
• Group comparisons require at least 5% contribution to total corpus.
• Below these thresholds, outputs are flagged as potentially noisy or incomplete.

All Level A processing is deterministic: identical input yields identical Level A metrics.

3) Level B: Narrative Synthesis

Level B uses a large language model to generate a reflective narrative summary of the selected user's communication patterns in group context.

Input to Level B includes Level A metrics plus contextual anonymised excerpts from the conversation.

Output sections include:

• Your role in the group (in context).
• Emotional tone and pressure points.
• Ideas, complexity, and conversational style.
• Values and moral framing (light touch).
• Practical upgrades (small, doable, non-cringy).
• Closing note.

Level B is interpretive, not deterministic. Re-running may produce different wording, while core pattern anchors remain tied to Level A data.

All speaker names except the user handle are anonymised in Level B output.

4) What this is not

• WhatYouSay is not a clinical, psychiatric, or diagnostic tool.
• It does not assess personality, mental health, intelligence, or character.
• Outputs describe language-use patterns and do not evaluate individuals.
• The Practical upgrades section offers observational suggestions, not therapeutic advice.

5) References

• Mohammad, S. M., and Turney, P. D. (2013). Crowdsourcing a Word-Emotion Association Lexicon. NRC Emotion Lexicon resource.
• Mohammad, S. M., and Turney, P. D. (2010). Emotions Evoked by Common Words and Phrases. Resource page and references.
• Araque, O., Gatti, L., and Kalimeri, K. (2022). LibertyMFD: A Lexicon to Assess the Moral Foundation of Liberty. DOI: 10.1145/3524458.3547264.
• Araque, O., Gatti, L., and Kalimeri, K. (2019). MoralStrength. arXiv:1904.08314.
• Vasse, BR. (2026). Measuring Within-Person Variation in Written Communication Patterns Across Social Contexts. Zenodo:18890804.