📝 Abstract:
Dialogue systems powered by large language models (LLMs) show strong generative abilities but often struggle with informal language, long-term coherence, and grounded responses in expert-driven conversations. This thesis presents three complementary methods to address these challenges. First, we introduce a novel temporal forecasting framework that models dialogue topic trajectories as time series and predicts future topic distributions based on Gaussian Mixture Model (GMM)–based similarity across conversations, offering interpretable and accurate forecasting in both domain-specific and open-domain settings. To enhance topic modeling under noisy user inputs, we develop SALDIRAY, a task-agnostic standardization pipeline that normalizes messages affected by spelling errors, slang, and abbreviations, improving downstream NLP performance across tasks such as sentiment analysis and topic labeling. Finally, we propose RAP (Retrieval-Augmented Paraphrasing), a retrieval-based generation method that retrieves similar past responses and paraphrases them using LLMs to fit the current context, significantly reducing hallucinations while maintaining stylistic alignment and relevance. Together, these methods improve the robustness, foresight, and factual grounding of conversational agents, advancing the reliability of LLM-based dialogue systems in real-world applications.