A Dialogue Response Generation Model with Latent Variable Modeling and LWE Attention based on the VW-BLEU Evaluation Metric

Abstract

In this page, We propose a novel dialogue response generation model that combines BART-large models with latent variable modeling and the Latent Weight Enhanced (LWE) Attention mechanism to improve the performance of dialogue response generation. Dialogue systems are increasingly utilized in various domains, such as customer service, medical consultations, and personal assistants. In these applications, generating natural and consistent responses is a critical challenge. However, existing models often struggle to maintain long-term dialogue coherence or capture hidden meanings and inferential relationships within conversations. To address these limitations, this research presents enhanced modeling techniques based on the Learning to Memorize Entailment and Discourse Relations (LMEDR) framework, effectively capturing contextual consistency and inferential relationships in dialogue. Additionally, a new evaluation metric, Vocabulary-Weighted BLEU (VW-BLEU), is introduced to overcome the limitations of traditional BLEU scores by assessing response diversity and naturalness. Experimental results demonstrate that the proposed model achieves superior performance on traditional metrics such as BLEU, METEOR, and ROUGE-L. In addition, our model exhibits higher lexical diversity as reflected by improved Distinct-1 (2.79 vs. 2.47) and Distinct-2 (14.95 vs. 13.82) scores and a significant boost in persona consistency with a Consistency Score of 39.41 compared to 25.31 from the state-of-the-art LMEDR. Furthermore, on the DSTC7-AVSD dataset, our model yields improvements of 13.4% in BLEU and 12.3% in METEOR, confirming its ability to generate responses with enhanced lexical richness and contextual fluency, as further validated by high VW-BLEU scores