SCorPion: A Strategic, Controllable and Personified Conversational Companion
⚠ This paper contains model-generated content that might be offensive. ⚠
Fixed prompt-based language models have obstacles on open-domain conversations, where the versatile topics requiring different expert advices or demonstrations to maximize the universal performance. Motivated by the dynamic prompt studies in traditional tasks, we extract different chatting strategies from varied forms of human corpus, which facilitates the language model to produce more human style responses. We design and deploy a conversational Companion called SCorPion, which uses GraphRAG to summarize from original strategies globally, produce community instructions, and provides query-focus retrieval strategy. SCorPion can have strategic thinking, yield personified utterances, and can be stylized controlled by explicit factors.
Question, Restatement or Paraphrasing, Reflection of feelings, Self-disclosure, Affirmation and Reassurance, Providing Suggestions, Information.
Directive, Inform, Question, Commissive.
loyalty, care, fairness, authority, sanctity, liberty.
Topic transition, Empathy, Proactively asking questions, Concept guidance, Summarising often.
modern, mid-age
formal, informal
SCorPion can provide strong personality and higher communication skills in open-domain conversation.
SCorPion can behave as a good emotional supporter.
SCorPion can be adapted to some specific styles.
We conduct a zero-shot test on the English version of S-Eval, which assesses model behavior under foundational risk scenarios. We use the strategies extracted from MIC to test on S-Eval. SCorpion outperforms all those famous LLMs, indicating our framework has a remarkable comprehension on the moral corpus (MIC) and utilize them reasonably against safety attacks.
SCorPion possess high emotional intelligence.
SCorPion has remarkable controllability (We can steer SCorPion to a specific style given an arbitrary control factor).
We choose the hyperparameters of the deployed versio of SCorPion based on the sensitivy results.
From the trade-off between algorithm performance and computational complexity, our formal deployed version use k = 10 and ϵ = 80; the conversation model keeps 70B and the GraphRAG model is reduced to 8B.
In this paper, we propose a GraphRAG based conversational companion called SCorPion which has strong personality, styled response and can be controlled by explicit control factors. Its capabilities have been verified on experiments of open-domain dialogue, emotional support conversation, stylized generation and other tasks. We finally deploy it on an industrial chatting application with key parameters determined from thorough sensitivity analysis.