Adaptive Intent-Preference Arbitration for Conversational Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Conversational recommender systems integrate long-term preferences (LTP) from previous sessions with short-term intent (STI) expressed in the current conversation. Fixed-rule fusion may produce suboptimal recommendations when these signals conflict. This paper proposes Adaptive Intent–Preference Arbitration (AIPA), which classifies their relationship into five categories: complement, consistent, conflict, override, and uncertain. Based on this classification, the model selects one of four actions: signal fusion, LTP prioritisation, STI prioritisation, or clarification. Candidates are re-ranked using learned, action-conditioned fusion weights. Additional components track recurring intent shifts across sessions and use counterfactual ablation analysis to identify the contribution of each signal. The framework is evaluated on ReDial, an English-language human–human movie recommendation corpus. Cross-session profiles are constructed from recurring seeker identities, while MovieLens genres provide item metadata. Since ReDial lacks relationship annotations, labels are generated using transparent weak-labeling rules and controlled synthetic injection covering all five categories. Results are reported separately for both sources. Training emphasises conflicting signals through conflict-weighted losses, label smoothing, and confidence-filtered self-training. All model-selection decisions use the validation split. Baselines are in-house reimplementations trained under a common protocol, limiting comparisons to this study. Across five random seeds using a pretrained sentence encoder, AIPA substantially outperforms profile-only, GRU, and SASRec baselines. However, its Hit@10 remains approximately one point below the intent-only and KBRD-style baselines across all turns and conflict subsets. Ablation results show negligible variation. The arbitration head achieves 0.99 clarification precision and a 1% wrong-override rate. Negative ranking results, an order-permutation control, MovieLens join coverage, and deployment measurements of peak memory and end-to-end latency are reported. Code, configuration files, and the executed notebook are released.