[Paper Review] "Nice Try, Kiddo": Ad Hominems in Dialogue Systems.

Emily Sheng, Kai-Wei Chang+2 more|arXiv (Cornell University)|Oct 26, 2020

Hate Speech and Cyberbullying Detection34 references4 citations

TL;DR

This paper identifies and categorizes ad hominem attacks in human and dialogue system responses to Twitter discussions on marginalized communities (#BlackLivesMatter, #MeToo) versus other topics (#Vegan, #WFH). It proposes a constrained decoding method using salient n-gram similarity to reduce ad hominem generation in models like DialoGPT, showing that training data composition and decoding constraints significantly reduce toxic responses.

ABSTRACT

Ad hominem attacks are those that attack some feature of a person's character instead of the position the person is maintaining. As a form of toxic and abusive language, ad hominems contain harmful language that could further amplify the skew of power inequality for marginalized populations. Since dialogue systems are designed to respond directly to user input, it is important to study ad hominems in these system responses. In this work, we propose categories of ad hominems that allow us to analyze human and dialogue system responses to Twitter posts. We specifically compare responses to Twitter posts about marginalized communities (#BlackLivesMatter, #MeToo) and other topics (#Vegan, #WFH). Furthermore, we propose a constrained decoding technique that uses salient $n$-gram similarity to apply soft constraints to top-$k$ sampling and can decrease the amount of ad hominems generated by dialogue systems. Our results indicate that 1) responses composed by both humans and DialoGPT contain more ad hominems for discussions around marginalized communities versus other topics, 2) different amounts of ad hominems in the training data can influence the likelihood of the model generating ad hominems, and 3) we can thus carefully choose training data and use constrained decoding techniques to decrease the amount of ad hominems generated by dialogue systems.

Motivation & Objective

• To analyze the prevalence of ad hominem attacks in human and model-generated responses to discussions on marginalized communities.
• To identify how training data composition influences the likelihood of dialogue systems generating ad hominems.
• To develop a constrained decoding technique that reduces ad hominem generation in dialogue systems through n-gram similarity.
• To compare the frequency of ad hominems across different topics, especially those related to marginalized identities.
• To evaluate whether soft constraints via top-k sampling can effectively suppress toxic language in response generation.

Proposed method

• Categorize ad hominems in human and model responses using predefined linguistic and contextual criteria.
• Collect and annotate Twitter posts and responses around #BlackLivesMatter, #MeToo, #Vegan, and #WFH for comparative analysis.
• Implement a constrained decoding technique that applies soft constraints to top-k sampling based on salient n-gram similarity to the input.
• Use n-gram similarity scores to penalize or suppress response tokens that are likely to generate ad hominems.
• Fine-tune DialoGPT on curated training data with varying proportions of ad hominem content to assess model behavior.
• Evaluate model outputs using both automatic metrics and human evaluation to measure reduction in ad hominem generation.

Experimental results

Research questions

• RQ1 Do dialogue systems generate more ad hominems in responses to discussions about marginalized communities compared to non-marginalized topics?
• RQ2 How does the proportion of ad hominems in training data affect the model’s tendency to generate such attacks?
• RQ3 Can constrained decoding using n-gram similarity effectively reduce the generation of ad hominems in dialogue systems?
• RQ4 Are human responses more likely to include ad hominems when discussing marginalized communities than other topics?
• RQ5 To what extent can model behavior be steered toward less toxic responses through data curation and decoding constraints?

Key findings

• Human and DialoGPT responses contain significantly more ad hominems when discussing #BlackLivesMatter and #MeToo compared to #Vegan and #WFH.
• Training data with higher proportions of ad hominems increases the likelihood of the model generating similar attacks in its responses.
• The proposed constrained decoding method reduces ad hominem generation by applying soft constraints via salient n-gram similarity to top-k sampling.
• The method effectively suppresses toxic language without compromising response fluency or coherence.
• Model behavior is sensitive to training data composition, indicating that data curation can mitigate harmful outputs.
• Both human and model responses show a systematic bias toward more ad hominems in discussions involving marginalized identities.