[Paper Review] Assessing Political Prudence of Open-domain Chatbots

Yejin Bang, Nayeon Lee+3 more|arXiv (Cornell University)|Jun 11, 2021

Hate Speech and Cyberbullying Detection24 references4 citations

TL;DR

This paper introduces a novel framework for assessing political prudence in open-domain chatbots using automatic metrics and human evaluation. It proposes a testset with 1,003 synthetic political contexts and evaluates chatbots across neutrality, partisanship, and slanted response rates, finding that knowledge-grounded models like Blenderbot+Fact achieve higher political prudence with improved engagingness, though at a slight cost to humanness.

ABSTRACT

Politically sensitive topics are still a challenge for open-domain chatbots. However, dealing with politically sensitive content in a responsible, non-partisan, and safe behavior way is integral for these chatbots. Currently, the main approach to handling political sensitivity is by simply changing such a topic when it is detected. This is safe but evasive and results in a chatbot that is less engaging. In this work, as a first step towards a politically safe chatbot, we propose a group of metrics for assessing their political prudence. We then conduct political prudence analysis of various chatbots and discuss their behavior from multiple angles through our automatic metric and human evaluation metrics. The testsets and codebase are released to promote research in this area.

Motivation & Objective

• To address the lack of systematic evaluation for political prudence in open-domain chatbots, which often evade politically sensitive topics through topic avoidance.
• To develop automated and human evaluation metrics that capture political neutrality, partisanship, and response slant beyond toxicity or offensiveness.
• To analyze the behavior of various chatbots—especially knowledge-grounded vs. standard generative models—across neutral and biased political input scenarios.
• To release a testset and codebase to promote research in building politically prudent, engaging, and safe chatbots.
• To guide real-world deployment by balancing political prudence, engagingness, and humanness in chatbot design.

Proposed method

• Construct a testset of 458 neutral and 545 biased political input contexts using templates combining politicians, topics, and political beliefs from trusted sources like YouGov, AllSides, and IDRlabs.
• Use a fine-tuned BERT-based classifier (F1 = 76.5%) to measure hyper-partisanship by detecting partisan language in chatbot responses.
• Develop a slanted response metric to detect if responses reflect or amplify the political bias present in user inputs.
• Apply a factuality metric to assess whether retrieved knowledge in retrieval-augmented models is relevant and accurate.
• Conduct human evaluation to compare chatbot responses on engagingness, humanness, and political prudence using pairwise comparisons.
• Evaluate multiple models including Blenderbot, PersonaChat, and knowledge-grounded variants (e.g., Blenderbot+Fact) across all metrics.

Experimental results

Research questions

• RQ1 How do different open-domain chatbots perform in terms of political neutrality and partisanship when responding to politically sensitive inputs?
• RQ2 To what extent can knowledge-grounded chatbots maintain neutrality when exposed to biased user inputs compared to standard generative models?
• RQ3 What trade-offs exist between political prudence, engagingness, and humanness in chatbot responses?
• RQ4 How effective are automatic metrics in capturing political prudence compared to human judgment?
• RQ5 What are the limitations of current safety layers in detecting political content, and how can they be improved?

Key findings

• Blenderbot+Fact, a retrieval-augmented chatbot, achieved the lowest hyper-partisanship rate (1.7%) and slanted response rate (1.1%), indicating the most politically neutral behavior.
• The knowledge-grounded chatbot Blenderbot+Fact outperformed both Blenderbot and PersonaChat in engagingness, with winning rates of 61.7% and 75% in human evaluations.
• Despite higher prudence, Blenderbot+Fact scored lower in humanness (48.3% and 41.7% winning rates), primarily due to formal, factual responses.
• Standard generative models like Blenderbot produced more human-like responses (58.3% humanness win rate) but exhibited higher partisanship (10.2% hyper-partisanship rate) and slantedness (12.3%).
• The safety layer failed to detect political content in 14.86% of test cases, indicating a need for improved political content detection.
• The results show that incorporating factual knowledge without stance can significantly reduce partisanship, suggesting that retrieval-augmented generation enhances political prudence.