Sentiment Analysis
WeCheck's Sentiment Analysis layer is the component of the AI engine responsible for understanding the meaning of what a subject has written, shared, or been associated with online — not just detecting the presence of certain words. It uses Natural Language Processing (NLP) models to analyze tone, context, narrative patterns, and behavioral signals across all discovered content.
Beyond Keyword Matching
Traditional screening tools work by flagging posts that contain specific keywords — a list of banned words or phrases. This approach fails in several predictable ways:
- Sarcasm and irony — "Oh great, another protest" is very different from "I'm organizing a protest"
- Context collapse — A post about "eliminating bugs in production" reads very differently from flagging actual threats
- Negation — "I do not support violence" contains the word "violence" but means the opposite
- Multilingual nuance — Idioms, slang, and culturally specific expressions don't translate literally
WeCheck's NLP models are trained to understand context and intent, not just surface-level keyword presence. A flag is generated only when the full context of a post or pattern of posts indicates a genuine signal.
What Gets Analyzed
The Sentiment Agent processes content across four dimensions:
1. Tone & Emotional Register
Is the subject's public writing consistently aggressive, inflammatory, or threatening? Or is a single post an anomaly in an otherwise neutral history? WeCheck evaluates patterns over time, not isolated incidents.
2. Thematic Content
Does the subject's public activity cluster around concerning thematic areas — extremist ideology, organized harassment campaigns, financial misconduct, or professional misrepresentation?
3. Narrative Consistency
Does the subject's stated professional history, claimed expertise, or personal narrative hold up across platforms? Inconsistencies in self-reported information are flagged as a risk signal.
4. Network Associations
Does the subject publicly engage with, endorse, or amplify content from known problematic actors or communities? Association signals are presented with full context, not as binary flags.
What Gets Flagged
WeCheck surfaces the following categories of behavioral signals as potential risk flags:
| Signal Category | Examples |
|---|---|
| Threatening or violent rhetoric | Direct or implied threats toward individuals or groups |
| Extremist content | Affiliation with or promotion of extremist ideologies |
| Hate speech | Targeted derogatory language based on protected characteristics |
| Professional misrepresentation | Claims that contradict verifiable employment or credential history |
| Financial misconduct signals | Public mentions of fraud, money laundering, or sanctions violations |
| Organized harassment | Participation in coordinated online harassment campaigns |
What Does NOT Get Flagged
WeCheck's models are specifically designed to avoid flagging content that is not relevant to professional risk assessment:
- Political opinions — Holding a political view, voting preference, or policy position is not a risk signal
- Religious expression — Religious beliefs, practices, or affiliations are not assessed
- Personal lifestyle — Relationship status, hobbies, dietary choices, and non-professional personal content are ignored
- Protected characteristics — Race, ethnicity, gender, sexual orientation, disability, and similar attributes are never used as risk factors
- Isolated anomalies — A single post taken out of context does not trigger a flag; the engine looks for patterns
This boundary is a deliberate design choice and is regularly reinforced through bias audits of the model outputs. See ISO 42001 & EU AI Act for details on WeCheck's compliance approach.
Multilingual Analysis
WeCheck's NLP models operate natively in 50+ languages without relying on machine translation as an intermediary step.
This matters because translation introduces errors — idioms break down, culturally specific insults lose meaning, and sarcasm rarely survives. By analyzing content in its original language, WeCheck maintains analytical accuracy for international investigations involving Spanish, Portuguese, Arabic, Mandarin, Russian, French, German, and dozens of other languages.
Confidence & Context in Reports
Sentiment findings in a WeCheck report are never presented as bare labels. Every flagged item includes:
- The source content — The specific post, comment, or publication that triggered the flag
- The platform and date — Where and when it was published
- The flag category — Which risk category it falls under (e.g., "threatening rhetoric")
- A confidence level — How strongly the model assessed this as a genuine signal vs. an ambiguous case
- Surrounding context — Adjacent content to give the reviewer the full picture before making a judgment
This design ensures that the human reviewer — not the algorithm — makes the final call on whether a flag is professionally relevant.