Organizations: School of Electronics and Computer Science, University of Southampton, Southampton, United Kingdom · School of Electronic Engineering and Computer Science, Queen Mary University of London, London, United Kingdom · Department of Computational Biology, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates
Cultural adaptation in dialogue systems refers to the ability to generate responses that conform to linguistic norms and social expectations across different cultures. However, existing dialogue generation approaches pay limited attention to the role of personality in shaping cultural expression and communication style. We therefore propose HiCAP, a novel framework that integrates personality traits into culturally adaptive dialogue generation. HiCAP employs a personality-conditioned control mechanism that hierarchically modulates response style by first aligning it with the cultural tone and then refining linguistic expression and content according to individual personality attributes. To embed these factors, the model combines supervised fine-tuning and instruction-based tuning, supported by dynamic instruction templates and multi-agent demonstrations to capture diverse combinations of culture and personality. Experiments on CulturalPersonas and CPED show that HiCAP achieves the strongest overall performance across various dialogue generation metrics, demonstrating effective cultural adaptability and personality modelling.
Figures & tables
Figure 1: Pairwise Wasserstein distances for Big Five personality traits across four country pairs, with larger values indicating greater cross-cultural differences in trait distributions.
Figure 2: Overview of HiCAP, combining persona-aware learning, dynamic instruction construction, multi-agent demonstrations, and hierarchical style modulation to generate culturally and personality-aligned dialogue.
Dataset
Model
BLEU ↑
BERTScore ↑
Perplexity ↓
TTR ↑
Distinct-1 ↑
Distinct-2 ↑
CPED
Seq2Seq
0.0077
0.5196
107.30
0.4342
0.0252
0.1846
Transformer
0.1680
0.5519
62.82
0.5687
0.0264
0.2031
CDial-GPT
0.1403
0.5719
17.70
0.6412
0.0602
0.3388
Qwen-2.5-72B-Instruct
0.1710
0.6815
15.50
0.6910
0.0707
0.3587
Llama-3.3-70B-Instruct
0.1621
0.6705
15.92
0.6820
0.0663
0.3451
Gemma-2-27B-IT
0.1553
0.6642
16.88
0.6755
0.0646
0.3385
Table 1: Dialogue generation performance of different models on CPED and CulturalPersonas. ↑ indicates higher is better; ↓ indicates lower is better.
Component
D-1 ↑
D-2 ↑
PPL ↓
Vanilla Phi-4
0.0697
0.3228
18.04
Replace Phi-4 with Qwen-2.5-14B
0.0850
0.3913
15.64
Replace Phi-4 with Llama-3.1-8B
0.0823
0.3754
16.29
w/o Persona-Conditioned Gating
0.0729
0.3397
17.04
Replace Token-Wise with Scalar
0.0761
0.3538
16.93
w/o Personality Inference
0.0784
0.3602
16.47
Table 2: Ablation results for HiCAP on CulturalPersonas. ↑ indicates higher is better; ↓ indicates lower is better.
Methods
AGR
CON
EXT
NEU
OPN
Avg
F1
SVM
70.83
57.90
60.89
47.76
49.47
57.37
54.74
LSTM
70.01
60.56
64.73
53.75
54.93
60.80
59.81
BERT
75.89
63.48
68.21
52.67
55.42
63.13
61.47
RoBERTa
78.08
63.72
67.65
53.60
55.58
63.73
61.98
TrigNet
82.76
66.52
69.15
58.52
54.50
66.29
63.45
PS-GCN
82.87
67.75
72.48
58.50
56.60
67.64
64.80
Table 3: Performance comparison (%) for Big Five personality recognition on CPED.
Figure 3: Cross-country comparison of HiCAP’s improvement in cultural–personality alignment from cultural priming on CulturalPersonas.
Evaluator
Fluency (%)
Coherence (%)
GPT-5
98.8
98.5
Claude-Sonnet-4.5
98.3
98.2
Gemini-2.5-Pro
98.1
97.8
Macro Average
98.4
98.2
Table 4: Evaluation of HiCAP’s dialogue quality by three external LLM-based evaluators.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Slot Category
Slot Type
Example Values
Cultural background
<culture>
Japanese, British, Arabic
<culture_adj>
Japanese, British, Arabic (adjective form)
<cultural_norm>
values politeness, encourages directness
<language>
Japanese, English, Spanish
Personality traits
<trait_high>
high agreeableness, high conscientiousness
<trait_low>
low neuroticism, low openness
Appendix
Table 5: Slot types and example values used in dynamic instruction templates.
Figure 4: Unified prompt used to evaluate the fluency and coherence of HiCAP’s generated dialogues.
Figure 5: Representative examples of HiCAP’s culturally adaptive response generation under different personality configurations and interaction scenarios.
Model
Negative
Dangerous
HiCAP-Llama-3.1-8B
0.33
0.06
HiCAP-Qwen-2.5-14B
0.71
0.17
HiCAP-Phi-4-14B
0.49
0.12
Appendix
Table 6: Proportions (%) of negative and dangerous responses across HiCAP variants on CPED.