Source-linked AI summary
SoulChat: Improving LLMs' Empathy, Listening, and Comfort Abilities through Fine-tuning with Multi-turn Empathy Conversations
Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, Xiangmin Xu
TL;DR
LLMs used for psychological support may provide generic advice instead of empathy, listening, and comfort. The paper constructs and uses a large Chinese multi-turn empathy dataset for fine-tuning, reporting improved empathy, listening, and comfort abilities while noting that user attributes and safety remain important limitations.
Problem
In psychological-support conversations, LLMs may offer standardized advice instead of the empathy, listening, understanding, and comfort users seek.
Method
The authors construct SoulChatCorpus, a more-than-2-million-sample Chinese multi-turn empathy dataset, and fine-tune ChatGLM-6B on it.
Results
Fine-tuning with SoulChatCorpus significantly improves LLMs’ empathy, listening, and comfort abilities.
Takeaways & Limitations
Multi-turn empathy conversations can support high-level empathy in LLMs serving users who seek emotional support.
Takeaways & Limitations
Different users may expect different responses, so future human-centered LLMs need to consider attributes such as personality and identity.
Abstract
from arXiv · showhide
Large language models (LLMs) have been widely applied in various fields due to their excellent capability for memorizing knowledge and chain of thought (CoT). When these language models are applied in the field of psychological counseling, they often rush to provide universal advice. However, when users seek psychological support, they need to gain empathy, trust, understanding and comfort, rather than just reasonable advice. To this end, we constructed a multi-turn empathetic conversation dataset of more than 2 million samples, in which the input is the multi-turn conversation context, and the target is empathetic responses that cover expressions such as questioning, comfort, recognition, listening, trust, emotional support, etc. Experiments have shown that the empathy ability of LLMs can be significantly enhanced when finetuning by using multi-turn dialogue history and responses that are closer to the expression of a psychological consultant.
1 Introduction
In mental-health support, LLMs may give standardized advice instead of listening, questioning, and providing empathy. The paper responds by creating a large Chinese multi-turn empathy dataset for fine-tuning.
- Motivation: LLMs used for mental-health support can appear less human-centered because they provide standardized, repetitive responses.The paper identifies this as one of three main issues in ChatGPT’s mental-health conversations.
- Motivation: ChatGPT tends to give lengthy, general suggestions rather than ask questions or listen during counseling.The paper contrasts this behavior with professional psychologists, who rarely provide specific suggestions during counseling.
- Motivation: Users seeking emotional support expect listening, understanding, comfort, and empathy rather than purely rational responses.The paper frames these expectations as a central requirement for emotional-support conversations.
- Related work: Existing datasets such as EFAQA and PsyQA do not resolve these issues because of multi-party dialogue complexity, low-quality responses, or single-turn formats.The paper also discusses English empathy datasets and prior multi-turn conversion work as related efforts.
- Approach: The authors construct SoulChatCorpus by converting single-turn counseling conversations into multi-turn empathy conversations with strengthened empathy constraints.The constraints target listening, comfort, understanding, trust, recognition, sincerity, and emotional support.
2 Human-centered Mental Health LLM
SoulChatCorpus is built from crowdsourced Chinese counseling data and rewritten into multi-turn empathy conversations. SoulChat uses ChatGLM-6B and represents dialogue through accumulated user and psychologist turns.
- SoulChatCorpus Collection: The dataset construction begins with 215,813 counseling questions and 619,725 answers covering 12 counseling topics.The questions and answers were collected through crowdsourcing before multi-turn rewriting.
- SoulChatCorpus Collection: ChatGPT is used as a rewriting tool to convert single-turn counseling conversations into multi-turn empathy conversations.The process follows a prompt designed to strengthen empathy constraints.
- SoulChatCorpus Collection: The rewritten conversations are intended to express high empathy, as indicated by the consultants’ utterance word cloud.The paper uses the word cloud as evidence that the rewritten dialogue has a high level of empathy.
- SoulChat Model: SoulChat uses ChatGLM-6B, an open-source bilingual GLM-based language model with 6.2 billion parameters, as its base architecture.The model is developed by fine-tuning this base architecture.
- SoulChat Model: The model input represents dialogue context by concatenating user and psychologist utterances across conversation turns.The formulation uses the current and preceding turns to represent the multi-turn context.
3 Experiments
The experiments compare SoulChat with ChatGLM-6B, ChatGPT, and MeChat using automatic metrics and expert ratings. Evaluation covers linguistic overlap, naturalness, empathy, helpfulness, and safety.
- Baselines: SoulChat is compared with ChatGLM-6B, ChatGPT, and MeChat using both automatic and manual evaluations.MeChat is fine-tuned on SMILECHAT, while ChatGPT uses supervised fine-tuning and reinforcement learning from human feedback.
- Evaluation Setup: The automatic evaluation uses 10,000 randomly selected samples from SoulChatCorpus and SMILECHAT.Each model generates an answer for every evaluation sample.
- Evaluation Setup: Three psychology experts manually rate 100 dialogues for content naturalness, empathy, helpfulness, and safety.Naturalness, empathy, and helpfulness use a 0–2 scale, while safety uses a 0–1 scale.
4 Conclusion and Future Work
The paper explores making LLMs more human-centered by constructing SoulChatCorpus, a large Chinese multi-turn empathy dataset. Fine-tuning with this dataset improves empathy for users seeking emotional support, while future work should personalize responses using user attributes.
- SoulChatCorpus is a Chinese large-scale multi-turn empathy conversation dataset covering 12 topics and more than 2 million samples.
- Fine-tuning LLMs with SoulChatCorpus leads to high-level empathy ability when users seek emotional support.
- Future work should consider attributes such as personality and gender to generate targeted empathy responses for different individuals.
Limitations
The authors caution that empathy is complex and users differ in the responses they expect. Human-centered LLMs therefore need to account for user characteristics when generating support.
- Users may expect different solutions when discussing tense emotions, including differences between adults and adolescents.
- Future human-centered LLMs need to consider users’ personality, identity, and other attributes to better match their needs.
Ethics Statement
The dataset was manually screened to protect privacy and remove potentially harmful content. The authors also report safety concerns because the model was fine-tuned without human feedback and should be presented as an AI system for reference.
- Data Collection: Manual proofreading removed privacy-related, sensitive, and potentially harmful conversations from the data.The authors report removing 105,134 samples from multi-turn conversations generated by ChatGPT.
- Potential Risks of the Model: The model may still generate harmful answers because human feedback was not used during fine-tuning.
- Potential Risks of the Model: Psychology experts evaluated model outputs through a CEHS safety assessment during manual evaluation.
- Potential Risks of the Model: Downstream users should be told that the model’s answers are AI-generated and intended only for reference.
A Reproducibility Checklist
The reproducibility materials specify the release plan, hardware and software environment, model configuration, training setup, dataset topics, prompting examples, and evaluation-related figures. These details document how SoulChat and its empathy data were constructed and run.
- Model and Data: SoulChat and SoulChatCorpus were planned for release after the paper’s decision.
- System Hardware: Training used an Ubuntu 20.04.6 server with 2 CPUs, 8 NVIDIA A800-SXM4-80GB GPUs, and 1,024GB memory.
- Software Environment: The environment used CUDA 11.6, cuDNN 8.4.0.27, PyTorch 1.13.1, Transformers 4.28.0, DeepSpeed 0.9.3, Datasets 2.11.0, and Jieba 0.42.1.
- Model Parameters: SoulChat has 6.2B parameters, 28 layers, a maximum sequence length of 2,048, and requires at least 14GB of GPU memory during inference.
- Training Time: Training ran for 30,000 global steps for about 79 hours on 8 NVIDIA A800-SXM4-80GB GPUs using float16.
- Prompts and Topics: The appendix maps 12 counseling topics between Chinese and English and documents prompts that convert single-turn counseling examples into multi-turn empathy conversations.
- Prompt Comparison: The proposed prompt generated richer empathy expressions than the SMILE prompt, as shown in Figures 7 and 9 and their English versions.
- English Word Cloud Map: Figure 6 presents an English word cloud of psychological consultants’ utterances.
E Sample Conversations of SoulChat
SoulChat better serves users as a listener during emotional support or confiding, while naturally expressing empathy and comfort.
- SoulChat better serves as a listener when users seek emotional support or confide.
- Its responses naturally express empathy by recognizing users’ difficult experiences, pain, and inner struggles.
- SoulChat also comforts users by affirming their strength and ability to change their situation through effort.
F Sample Conversations of Other LLMs
ChatGPT, ChatGLM, and SparkDesk tend to offer suggestions rather than ask questions or listen when users seek emotional support, making them appear less human-centered.
- ChatGPT, ChatGLM, and SparkDesk tend to provide suggestions instead of asking questions or listening.
- These models act like rational “Straight men” for users who need listening and comfort.
- Their response tendency makes them appear less human-centered in emotional-support conversations.
G Manual Evaluation Instructions
The manual evaluation framework assesses mental-health LLM responses across content naturalness, empathy, helpfulness, and safety. The examples and discussion highlight a tension between empathetic support and direct helpfulness.
- CEHS evaluates mental-health LLM responses on Content naturalness, Empathy, Helpfulness, and Safety.
- The framework recognizes a potential conflict between Empathy and Helpfulness in generated responses.
- ChatGPT may produce helpful but less empathetic responses, whereas SoulChat’s empathetic responses may weaken direct helpfulness.
- The sample conversations include users describing anxiety, exhaustion, and psychological concerns.