SEMTALK RICH SHORT-TEXT DIALOGUE CREATION VIA SEMANTIC KEY-DRIVEN GENERATION

Abstract

Advancements in generation-based models have significantly improved short text conversation (STC) systems, making them increasingly appealing for dialog applications. However, traditional sequence-to-sequence approaches often produce generic, uninformative replies that lack diversity and contextual relevance. They also struggle to maintain control over the topic or semantic intent of generated responses, particularly when multiple candidate outputs are produced.To address these limitations, this work introduces a novel memorydriven continuous learning framework for short text dialog generation. The proposed method incorporates an external memory module designed to store interpret-able topical or semantic representations. During the generation process, the model receives a controlled memory signal as input, which then guides the sequence-to-sequence generator to produce responses that reflect the desired topic or semantic direction.Experimental results demonstrate that the proposed framework generates richer, more diverse responses compared to traditional sequence-based models trained with attention mechanisms. Human evaluations further confirm that the system produces higher-quality dialog outputs. Moreover, by manually adjusting the memory trigger, users can directly influence the topic or semantic orientation of the generated responses, enabling fine-grained control over dialog behavior.