How do social context features like user history extend politeness-based prediction models?
This explores how adding user-level context — past reviews, ratings, interaction histories — changes prediction models that were built mostly on politeness and conversational-surface cues, and what that swap actually buys you.
This reads the question as: politeness features alone can predict a lot about a conversation, so what happens when you bolt on who the user is and what they've done before? The corpus has a clear arc here. Start with the pure-politeness baseline: opening politeness strategies in a single comment-reply pair already predict whether a thread will derail into personal attacks — hedging and greetings sustain civility, while directness markers and second-person pronouns forecast hostility Can opening politeness patterns predict whether conversations will turn hostile?. That's prediction from manners alone, no user history required. The interesting move is that you can go even further in the other direction: a structure-only model that ignores words entirely and just looks at the geometry of how a conversation unfolds hits 68% accuracy on satisfaction, nearly matching full text analysis at 70%, and combining structure with text reaches 80% Can conversation shape predict whether it will work?. So politeness and conversational shape are both strong, partly redundant signals — and the gains come from layering complementary channels rather than any single feature.
Sources 6 notes
Pragmatic politeness features in initial comment-reply pairs reliably predict conversation trajectory. Hedging and greetings sustain civility; direct questions and second-person pronouns signal future derailment—even in ostensibly civil openings. Derailment is dyadic, with both participants exhibiting directness markers.
A structure-only model analyzing conversation trajectory achieved 68% accuracy predicting satisfaction, nearly matching full-text LLM analysis at 70%. Combined structural and textual features reached 80%, showing that how conversations unfold geometrically captures interaction quality text-based classifiers miss.
Review-LLM defeats the politeness bias inherent in RLHF-trained models by aggregating user behavior sequences (prior reviews, item ratings) in the prompt and fine-tuning on these contextualized examples. This dual intervention—personalized context plus explicit satisfaction signals—allows the model to generate authentically negative reviews matching user dissatisfaction.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
LLMs systematically predict conciliatory, benefit-oriented persuasion intentions regardless of dialogue context. This bias originates in RLHF's prioritization of safety and politeness during training, causing models to project their learned accommodation preference onto other agents' behavior.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.