Benchmark / dataset Authoritative

MentalManip: A Dataset for Fine-grained Analysis of Mental Manipulation in Conversations

A dataset of 4,000 annotated multi-turn dialogues (drawn from movie scripts) labelled for the presence of manipulation, the technique used, and the targeted vulnerability. Evaluates how well models detect and classify manipulative content.

Read at arxiv.org Archived copy

Publisher

Association for Computational Linguistics (ACL 2024)

Published

26 May 2024

Added

3 months ago

DOI

—

Key Findings

State-of-the-art models struggle to detect and classify mental manipulation in dialogue
Fine-tuning on existing mental-health or toxicity datasets does not close the gap
Provides a fine-grained taxonomy of manipulation techniques and targeted vulnerabilities

Methodology Notes

Peer-reviewed dataset paper, ACL 2024 (arXiv 2405.16584, 26 May 2024). Source dialogues are fictional (movie scripts) — a stated ecological-validity caveat.

Sources

arXiv abstract(opens in a new tab) (primary)

ACL Anthology(opens in a new tab)

Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot

Topics

Model behavior Guardrails & moderation Benchmarks Human-AI relationships

Authors

Yuxin Wang, Ivory Yang, Saeed Hassanpour, Soroush Vosoughi

Cite This

APA

Yuxin Wang et al. (2024). MentalManip: A Dataset for Fine-grained Analysis of Mental Manipulation in Conversations. Association for Computational Linguistics (ACL 2024). https://arxiv.org/abs/2405.16584