This paper presents PIDE, a novel neural architecture that jointly models long-term user preferences through historical interaction encoding and short-term intent evolution via session context modeling. PIDE introduces a dual-path transformer architecture, a disambiguation-aware attention mechanism, and a personalized query representation layer. Evaluation on conversational search benchmarks shows improvements in intent accuracy over state-of-the-art baselines.
Key findings
PIDE jointly models user history and session context for intent disambiguation.
The dual-path architecture enables cross-attention between user history and session context.
Disambiguation-aware attention identifies and separates multiple intents within ambiguous queries.
Personalized query representation fuses user-specific and session-specific signals.
PIDE outperforms state-of-the-art baselines in conversational search benchmarks.
Limitations & open questions
The paper does not discuss the scalability of PIDE for very large user histories.
The effectiveness of PIDE in diverse languages and cultures is not evaluated.