Abstract:Recently, there has been a heightened interest in building chatbots based on Large Language Models (LLMs) to emulate human-like qualities in multi-turn conversations. Despite having access to commonsense knowledge to better understand the psychological aspects and causality of dialogue context, even these powerful LLMs struggle to achieve the goals of empathy and emotional support. Current commonsense knowledge derived from dialogue contexts is inherently limited and often fails to adequately anticipate the future course of a dialogue. This lack of foresight can mislead LLMs and hinder their ability to provide effective support. In response to this challenge, we present an innovative framework named Sensible and Visionary Commonsense Knowledge (Sibyl). Designed to concentrate on the immediately succeeding dialogue, this paradigm equips LLMs with the capability to uncover the implicit requirements of the conversation, aiming to elicit more empathetic responses. Experimental results demonstrate that incorporating our paradigm for acquiring commonsense knowledge into LLMs comprehensively enhances the quality of their responses.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to address the challenges that current large - language models (LLMs) encounter when generating empathetic and emotionally supportive responses in multi - turn conversations. Although these powerful LLMs can access common - sense knowledge to better understand the psychological aspects and causal relationships of conversations, they still struggle to achieve the goals of empathy and support. Specifically: 1. **Limitations of existing common - sense reasoning methods**: - The common - sense knowledge currently derived from the conversation context has inherent limitations and cannot fully anticipate the future direction of the conversation. - This lack of foresight may mislead LLMs and impede their ability to provide effective support. 2. **Insufficiency in predicting the future of conversations**: - Existing methods mainly rely on conversation history and overlook the possibilities of the future of conversations. Due to the one - to - many nature of conversation generation, common - sense knowledge based on conversation history may not be sufficient to generate ideal responses. 3. **Diverse response requirements**: - In a given conversation context, there are multiple possible appropriate responses, each associated with different conversation common - sense. Methods based solely on conversation history are prone to introducing noisy information, causing the language model to overlook the needs for empathy and support. To solve these problems, the author proposes a new paradigm - **Sensible and Visionary Commonsense Knowledge (Sibyl)**, which enables LLMs to identify potential causal factors, the psychological states of participants and their possible intentions by dynamically deriving common - sense knowledge related to the future of conversations, thereby generating more empathetic and supportive responses. ### Main contributions of Sibyl 1. **Addressing the limitations of existing common - sense reasoning methods**: - It resolves the confusion of common - sense knowledge caused by the one - to - many problem and prevents the language model from ignoring the goals of empathy and support. 2. **Proposing an innovative multi - turn conversation common - sense reasoning paradigm**: - It combines psychological, emotional and causal factors to derive common - sense knowledge related to the future of conversations. 3. **Experimental verification of effectiveness**: - Extensive experiments show that Sibyl significantly improves the ability of LLMs to generate high - quality responses in multiple settings and performs well in both automated and human evaluations. Through these improvements, Sibyl can help LLMs handle complex conversation situations more effectively and promote the generation of empathy and emotional support.

Sibyl: Empowering Empathetic Dialogue Generation in Large Language Models via Sensible and Visionary Commonsense Inference

Rational Sensibility: LLM Enhanced Empathetic Response Generation Guided by Self-presentation Theory

Harnessing the Power of Large Language Models for Empathetic Response Generation: Empirical Investigations and Improvements

Sibyl: Simple yet Effective Agent Framework for Complex Real-world Reasoning

Improving Empathetic Dialogue Generation by Dynamically Infusing Commonsense Knowledge.

Think Before You Speak: Cultivating Communication Skills of Large Language Models via Inner Monologue

Building Emotional Support Chatbots in the Era of LLMs

Harnessing Large Language Models' Empathetic Response Generation Capabilities for Online Mental Health Counselling Support

Talk With Human-like Agents: Empathetic Dialogue Through Perceptible Acoustic Reception and Reaction

Supervised Knowledge Makes Large Language Models Better In-context Learners

Can Large Language Models Be Good Companions? An LLM-Based Eyewear System with Conversational Common Ground

Towards Joint Modeling of Dialogue Response and Speech Synthesis based on Large Language Model

A New Dialogue Response Generation Agent for Large Language Models by Asking Questions to Detect User's Intentions

WisdomBot: Tuning Large Language Models with Artificial Intelligence Knowledge

SoulChat: Improving LLMs' Empathy, Listening, and Comfort Abilities through Fine-tuning with Multi-turn Empathy Conversations

Empathetic Dialogue Generation via Sensitive Emotion Recognition and Sensible Knowledge Selection

Augmenting End-to-End Dialogue Systems with Commonsense Knowledge

Reasoning in Conversation: Solving Subjective Tasks through Dialogue Simulation for Large Language Models

Large Language Model based Situational Dialogues for Second Language Learning

CAB: Empathetic Dialogue Generation with Cognition, Affection and Behavior

Enhancing Empathetic Response Generation by Augmenting LLMs with Small-scale Empathetic Models