PRCCF: A Persona-guided Retrieval and Causal-aware Cognitive Filtering Framework for Emotional Support Conversation
Authors: Yanxin Luo, Xiaoyu Zhang, Jing Li, Yan Gao, Donghong Han
Organizations: School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China · Oracle Cloud Infrastructure, Oracle, Santa Clara, CA, United States
Emotional Support Conversation (ESC) aims to alleviate individual emotional distress by generating empathetic responses. However, existing methods face challenges in effectively supporting deep contextual understanding. To address this issue, we propose PRCCF, a Persona-guided Retrieval and Causality-aware Cognitive Filtering framework. Specifically, the framework incorporates a persona-guided retrieval mechanism that jointly models semantic compatibility and persona alignment to enhance response generation. Furthermore, it employs a causality-aware cognitive filtering module to prioritize causally relevant external knowledge, thereby improving contextual cognitive understanding for emotional reasoning. Extensive experiments on the ESConv dataset demonstrate that PRCCF outperforms state-of-the-art baselines on both automatic metrics and human evaluations. Our code is publicly available at: https://github.com/YancyLyx/PRCCF.
Figures & tables
Figure 1: An example illustrating how PRCCF generates emotional support responses. Persona-aligned demonstrations are retrieved to match the user’s expressive style, while causal cues and filtered commonsense knowledge are integrated as cognitive information to guide empathetic response generation.
Figure 2: The figure illustrates PRCCF framework, consisting of three main modules: the Persona-guided Multi-View Retriever, the Causality-aware Cognitive Filtering, and the Multi-Source Fusion for Generation.
Model
ACC(%) ↑
PPL ↓
B-1 ↑
B-2 ↑
B-3 ↑
B-4 ↑
D-1 ↑
D-2 ↑
R-L ↑
MIME (2020 EMNLP) Majumder et al. (2020)
-
43.27
16.15
4.82
1.79
1.03
2.56
12.33
14.83
BlenderBot-Joint (2021 ACL) Liu et al. (2021)
27.72
18.11
16.99
6.18
2.95
1.66
3.27
20.87
15.13
MISC (2022 ACL) Tu et al. (2022)
31.63
16.16
-
7.31
-
2.20
4.41
19.71
17.91
GLHG (2022 IJCAI) Peng et al. (2022b)
-
15.67
19.66
7.57
3.74
2.13
3.50
21.26
16.37
KEMI (2023 ACL) Deng et al. (2023)
-
15.92
-
8.31
-
2.51
-
-
17.05
FADO (2023 Knowledge-Based Systems) Peng et al. (2023)
32.41
15.52
-
8.31
4.36
2.66
3.80
21.39
18.09
Table 1: Results of automatic evaluation on ESConv, where the values of baselines are sourced from public papers. The bolded and underlined results indicate best and second-best result respectively.
Model
B-2 ↑
R-L ↑
ChatGPT 1-shot Zhao et al. (2023a)
4.53
13.19
ChatGLM-6B w/P-Tuning Bao et al. (2024)
7.22
16.15
ChatGPT w/Example Expansion Kang et al. (2024)
7.45
15.22
FSM Zhao et al. (2025)
3.25
15.80
straQ ∗ -distill Wang et al. (2025)
6.25
15.44
PRCCF
10.71
19.78
Table 2: Comparisons with Large Language Models, where the values of large language model baselines are sourced from public papers.
Model
PPL ↓
B-1 ↑
B-2 ↑
B-3 ↑
B-4 ↑
R-L ↑
PRCCF
13.10
23.94
10.71
5.87
3.55
19.78
w/o PR
13.19
17.27
6.95
3.53
2.05
16.76
w/o Per sim
13.28
22.89
9.96
5.37
3.22
19.13
w/o CCF
14.76
20.49
8.72
4.59
2.71
17.63
w/o Causal
13.27
22.16
9.59
5.13
3.05
19.01
w/o Filter
14.59
21.70
9.60
5.18
3.10
18.07
Table 3: The results of ablation study for PRCCF.
Figure 3: Top- n accuracy comparison across different methods.
Figure 4: Effect of the number of retrieved candidate pairs ( pairs ) in the PR module. All metrics are min–max normalized, with perplexity inverted so that higher values indicate better fluency. Performance peaks at pairs=5 .
Comparisons
Aspects
Win
Lose
Tie
Ours vs. D 2 RCU
Identification
58 ‡
32
10
Comforting
60 ‡
33
7
Suggestion
47
36
17
Overall
57 ‡
37
6
Ours vs. DKPE
Identification
47
36
17
Comforting
50 †
37
13
Table 4: The results of the human evaluation (%). † and ‡ indicate improvements with p-value <0.1 and p-value <0.05 , respectively. Significance is tested using a one-sided exact sign test on non-tied cases (ties are excluded).
Figure 5: Parameter update dynamics of the Cognitive Knowledge Refinement across training epochs.
Situation
My job just laid off 10% of the workforce and I’m afraid I’m next.
Context
Seeker
My company just laid off 10% of the workforce and furloughed another 10%. I’m worried I’m going to be in the next round.
Supporter
Why do you think you are the next round? Do you know how did they choose them?
Seeker
Just based on their tenure and job experience. I’m around the same level.
Supporter
I know, this year was stressful for everyone in our planet, so you are not alone…
Table 5: Case study of the generated responses by baselines and PRCCF.
Figure 6: The strategy distribution in the different stage of conversation.
Emotional support conversation studies how people and AI systems provide empathetic, safe help through dialogue, including counseling, peer support, and companionship.