TL;DR
TaskBridge is a "catalog-anchored" framework designed to bridge the lexical gap between high-level user goals (e.g., "make lasagna") and specific product titles in e-commerce catalogs. By grounding LLM-based query reformulations in actual catalog vocabulary and selecting diverse candidates, it secured first place in the SIGIR 2026 eCommerce Challenge.
What Problem Does This Paper Address?
Standard e-commerce search engines often rely on lexical matching (like BM25), which works well when a user searches for "Nike running shoes" but fails when they search for a task like "complete home office makeover." These task-oriented queries describe activities or goals rather than specific products, creating a massive "lexical gap."
The challenge is twofold:
- Under-reformulation: The query remains too vague to match specific product titles or categories.
- Over-reformulation: An LLM might suggest products that sound plausible but aren't actually in the catalog, or it might drift away from the user's original intent.
The authors aim to transform these abstract goals into searchable, catalog-grounded expressions while strictly preserving the original intent.
Methods
TaskBridge is an inference-time system (no training required) that uses a three-stage pipeline:
- Catalog Semantic Anchoring: The system extracts 2-5 gram phrases from the product corpus to build a phrase lexicon. It then uses Qwen3.5-9B to "purify" these into semantic anchors—normalized expressions of purchasable categories, objects, or accessories.
- Unified Five-Candidate Rewriting: Instead of generating one "best" rewrite, TaskBridge generates five distinct candidates representing different hypotheses:
q_raw: The original query (to protect against bad rewrites).q_fb: Uses multi-round retrieval feedback to iteratively refine the query.q_vote: A multi-role intent model where three LLM "personas" (Conservative, Taxonomy, Intent) vote on whether and how to add catalog terms.q_need&q_prod: Lexicon-constrained routes that select phrases directly from the extracted catalog resources.- Protocol-Aware Selection:
- Automatic Track: Selects the single candidate with the highest "proxy quality score" (calculated by reranking the top 10 BM25 results using Qwen3-VL-Reranker-8B).
- Interactive Track: Selects up to four candidates using Pareto selection, balancing high quality with low result overlap (diversity) to cover more potential product needs.
Datasets
The system was evaluated on the SIGIR 2026 eCommerce Challenge benchmark. This includes:
- A large-scale product-title corpus.
- Task-oriented queries for both an Open Phase (public leaderboard) and a Closed Phase (final evaluation on unseen data).
Key Insights and Ideas
- Catalog Grounding is Essential: Generative models are powerful, but they are prone to "hallucinating" product terms that don't exist in a specific store's inventory. By anchoring the LLM to a pre-extracted phrase lexicon, TaskBridge ensures the reformulations are actually "retrievable."
- The Value of Diversity: In the Interactive track, the authors found that providing four complementary queries (e.g., one focusing on furniture, another on lighting) is more effective than providing four variations of the "best" query.
- Multi-Role Consensus: Using different role prompts (Taxonomy vs. Intent) helps mitigate "query drift." If the different roles don't agree that a reformulation is necessary, the system defaults to the original query.
Key Results
TaskBridge achieved first place in the challenge. Quantitative highlights include:
- Open Phase: Improved NDCG@10 from 0.42 (baseline) to 0.49 in the Automatic track and 0.58 in the Interactive track.
- Closed Phase: Achieved a Task Completion NDCG@10 of 0.27 (Automatic) and 0.31 (Interactive).
- Interactive Gains: The Interactive protocol consistently outperformed the Automatic protocol, proving that multiple reformulation directions better capture the multifaceted nature of task-oriented searches.
Limitations and Future Work
- Model Dependency: The system relies heavily on the Qwen model family for generation, embedding, and reranking. Performance might vary significantly with different base models.
- Iterative Noise: While multi-round feedback (
q_fb) is powerful, the authors note that robustness to noisy signals during iterative refinement remains a challenge. - Lexical Focus: The final retrieval stage is still purely lexical (BM25). Future work could explore integrating dense retrieval or hybrid approaches while maintaining the strict catalog grounding that made TaskBridge successful.
Summary updated