MultiRECLAIM Annotation Platform & Reclamation-Aware Augmentation (RAA)

Community-based annotation system for LGBTQ+ slur reclamation research.

Live platform: https://multireclaim.study · https://annotate.multireclaim.study
Contact: contact@multireclaim.study

---

Overview

• MultiRECLAIM is a self-hosted, privacy-first annotation platform that collects community judgements on slur reclamation from the LGBTQ+ community and allies. It implements an asymmetric annotation schema based on the FATA framework, and zero-trust annotator privacy.

• Reclamation-Aware Augmentation: Synthetic contrastive pair generation. For every accepted reclaimed example the pipeline generates a matching non-reclaimed counterpart using the same slur.

---

Motivation

A fundamental challenge in the automated detection of marginalized language is the scarcity of implicit, naturalistic data. Building upon the foundational work of Ferrando et al. (2026) in the MultiPRIDE shared task, this thesis addresses a recognized structural challenge in slur reclamation datasets: distributional skew.

Reclamation-Aware Augmentation (RAA) Architecture

StanceConfig (slur, speaker, pragmatic_role, community, language)
    │
    ▼  Stage 1 — Base model generation (no RLHF)
    ▼  Stage 2 — Instruct validator (reclaimed only)
    ▼  Stage 3 — MPEC gate  (empowerment factors)
    ▼  Stage 4 — Feedback-driven regeneration (max attempts)
    │
    └─ Counterpart: base model + surface-mirroring prompt + MPEC

MultiRECLAIM Annotation Platform Architecture

Five Docker Compose services:

Internet
    │
    ▼
 Nginx
    ├── multireclaim.study        → Registration
    └── annotate.multireclaim.study → Label Studio
                                           │
                                       PostgreSQL
                                           │
                                        Cron
                                   (sync + export + backup)

Service	Image	Role
postgres	postgres:15-alpine	Annotation database
labelstudio	heartexlabs/label-studio:latest	Annotation interface
registration	custom FastAPI	Annotator intake + privacy gate
nginx	nginx:alpine	Reverse proxy + SSL termination
cron	python:3.11-slim	Nightly automation

---

Annotation Schema

Asymmetric 4-step schema — Text A (reclaimed candidate) gets Steps 1–3, Text B (generated counterpart) gets Steps 4a–4b only.

---

Project Configuration

434 raw pairs generated across 11 English slurs from MultiPRIDE Table 7 (fag, bitch, gay, puss, queer, drag, whore, queen, slut, sissy, slay), 40 target pairs per slur, on A40 GPU, bf16. After deduplication: 369 clean pairs, and after inverted removal 337 final pairs.

Project	ID	Tasks	Overlap	Purpose
Calibration Task	4	5	999	All annotators — schema familiarisation
Main Task	5	332	3	Research dataset — majority vote

Calibration overlap=999 means every annotator sees all 5 calibration tasks regardless of how many others have completed them. Main overlap=3 means each task needs exactly 3 independent annotations before closing.

---

Privacy Architecture

• No PII collected: names, real emails, IP addresses, device info — never stored
• Anonymous IDs: annotator_XXXXXX hash — no reverse lookup possible
• Generated credentials: annotator_XXXXXX@multireclaim.study email + random password, shown once
• Registry stores only: hash ID + gate responses + timestamp
• Zero-trust design: even if the registry is leaked, no individual is identifiable
• Session isolation: annotator sessions cleared on org assignment to prevent stale state

---

Security

• UFW: only ports 22, 80, 443 open
• Services bound to 127.0.0.1 (Docker/UFW bypass prevention)
• Nginx: default_server returns 444 for unknown hosts
• Nginx: rate limiting on registration (5r/m) and annotation (30r/m)
• Nginx: token injection on /user/signup/ for Label Studio DISABLE_SIGNUP_WITHOUT_LINK
• HSTS, X-Frame-Options, X-Content-Type-Options, Referrer-Policy headers
• SSH: key-only authentication
• Fail2ban: ban after 5 failed attempts
• Unattended-upgrades: automated OS security patches

---