A social information sensitive model for conversational recommender systems.

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Information Retrieval

Background:

• Conversational recommender systems (CRS) enable natural language interactions for item suggestions.
• Integrating conversational history with user social information remains a challenge for semantic fusion.
• Existing methods struggle to effectively bridge the semantic gap between diverse data sources.

Purpose of the Study:

• To develop an innovative framework for extracting and utilizing social information within conversational datasets.
• To introduce a novel social information sensitive semantic fusion (SISSF) method.
• To enhance the personalization and relevance of recommendations in CRS by incorporating social context.

Main Methods:

• Extracting social information by inferring ratings and constructing user-item/user-user interaction graphs.
• Employing contrastive learning (CL) within the SISSF method to bridge semantic gaps.
• Evaluating the framework on ReDial and INSPIRED datasets using automatic and human metrics.

Main Results:

• SISSF demonstrated significant improvements over baseline models on both datasets.
• Achieved superior performance in recommendation tasks (e.g., R@1, R@50) and conversational quality (e.g., Distinct metrics).
• Human evaluations confirmed marked improvements in fluency and informativeness.

Conclusions:

• The proposed SISSF framework effectively integrates social context into CRS.
• Contrastive learning is a viable approach for semantic fusion in conversational recommendation.
• Incorporating social context significantly enhances personalization and relevance in conversational systems.