Abstract:
This thesis presents a novel enhancement to conversational recommender systems
through the strategic integration of an advanced intent detection module using BART
(Bidirectional and Auto-Regressive Transformers). This integration builds upon an existing
Coarse-to-Fine Contrastive Learning framework initially introduced by previous
research. The primary innovation of this work is the development and deployment of
a robust intent detection module designed to enhance system understanding and user
interaction within conversational settings.
The intent detection module utilizes BART for the precise classification and labeling
of user intents extracted from dialogues. The system’s capacity to generate highly customized
recommendations is enhanced by this accurate categorization, which enables
the system to recognize and comprehend the complex needs of users. By analyzing user
dialogues and their corresponding intents, the module ensures that the system adapts
to each individual’s unique preferences and interaction patterns.
In addition to dialogue analysis, this research innovates by integrating intent-labeled
dialogues derived from user reviews into the existing learning process. This integration
is crucial for refining user profiles and enhancing the granularity with which the system
understands and predicts user behavior. The enhanced model leverages these labeled
dialogues to feed into both coarse and fine-grained stages of the contrastive learning
process, thereby improving the overall recommendation accuracy and user satisfaction.
Experimental results validate the effectiveness of integrating BART-based intent detection
into the conversational recommender system. Tests demonstrate that this method
significantly enhances the relevance and personalization of recommendations, outperforming
traditional models in conversational settings. This advanced intent detection
technique enables dynamic adaptation to user activities, that is important advancement
of recommender systems’ development, making certain that suggestions are appropriate
for the user and that they are accurate in their context.
