Source-linked AI summary
TaskBridge: Technical Report for SIGIR 2026 eCommerce Challenge
Zhiheng Fu, Yupeng Hu, Zixu Li, Zhiwei Chen, Qinlei Huang, Shaokun Wang, Han Liu, Fan Liu, Weili Guan, Liqiang Nie
TL;DR
Task-oriented e-commerce queries often imply product needs without using catalog language, creating a gap for lexical retrieval. TaskBridge grounds reformulations in catalog vocabulary and selects queries for the Automatic or Interactive protocol. The paper reports higher scores for Interactive than Automatic in both evaluation phases, suggesting complementary reformulations can better support retrieval.
Problem
Task-oriented queries express product needs through activities or outcomes, often without the product terms BM25 matches in catalog titles.
Source: §1 · p. 1Method
TaskBridge builds catalog phrase resources, generates complementary candidate reformulations, and selects a query set according to the retrieval protocol.
Source: §1 · p. 2, §2 · p. 2, §2.4 · p. 5Results
Interactive achieves higher scores than Automatic in both evaluation phases.
Source: §4 · p. 6Takeaways & Limitations
The conclusion suggests that complementary reformulations can better support task-oriented product retrieval.
Source: §4 · p. 6The problem definition treats preservation of task intent as an objective to pursue “as much as possible” while narrowing the lexical gap.
Source: §2.1 · p. 2
Abstract
from arXiv · showhide
The SIGIR 2026 eCommerce Challenge Query Reformulation Task evaluates the reformulation of task-oriented queries into expressions suitable for product retrieval, followed by BM25-based retrieval. These queries often imply product needs through activity-level semantics, creating a lexical gap with catalog descriptions. Systems must preserve task intent while adapting the number of reformulations to the Automatic and Interactive protocols. We propose TaskBridge, a catalog-anchored framework comprising Catalog Semantic Anchoring, Unified Five-Candidate Rewriting, and Protocol-Aware Selection. It grounds queries in catalog vocabulary, generates five complementary reformulations, and selects one query for Automatic retrieval or up to four for Interactive retrieval using BM25 scores, query-product reranking, and weighted Jaccard overlap. The selected queries are independently executed using BM25 to produce results in the required six-column format. Experiments demonstrate improved retrieval effectiveness, with multiple complementary queries providing additional gains in the Interactive setting. TaskBridge achieved first place in the SIGIR 2026 eCommerce Challenge.
1 Introduction
Task-oriented e-commerce queries express product needs through activities or outcomes, often without matching catalog terms. TaskBridge addresses this gap by grounding reformulations in catalog vocabulary while preserving task intent and adapting selection to the submission protocol.
1 Introduction
- Task-oriented reformulation must make implicit product needs explicit in catalog-matching language without changing the task meaning [1– 7].Queries such as “make lasagna” describe a task without naming the products needed to complete it.Source: §1 · p. 1
- Under-reformulation can miss required product categories, while over-reformulation can add unnecessary products and weaken intent.Language-model-plausible terms alone do not guarantee a match to catalog titles or categories.Source: §1 · p. 1
- Because BM25 ranks through term matches to product titles, it cannot infer required products when task queries lack catalog product terms [24, 25].The challenge is a controllable mapping between task semantics and catalog vocabulary [8, 9].Source: §1 · pp. 1–2
- TaskBridge builds catalog phrase resources, generates candidate reformulations, selects queries for the protocol, and submits them to BM25 retrieval.Its framework combines Catalog Semantic Anchoring, Unified Five-Candidate Rewriting, and Protocol-Aware Selection.Source: §1 · p. 2
- The SIGIR eCommerce Query Reformulation Task evaluates this catalog-grounding challenge, and the paper validates TaskBridge on its benchmark.The introduction frames the task around preserving intent while matching catalog expressions.Source: §1 · pp. 1–2
2 Method
TaskBridge maps task-oriented queries into catalog-oriented reformulations while preserving task intent, then selects queries according to the retrieval protocol. It combines catalog phrase resources, complementary candidate generation, quality and diversity scoring, and independent BM25 retrieval.
2 Method
- TaskBridge is an inference-time framework that organizes catalog resources, candidate generation, and protocol-aware selection without model training.It uses Qwen3.5-9B for generation and judgments, Qwen3-VL-Embedding-8B for anchor retrieval, and Qwen3-VL-Reranker-8B for selection scoring.Source: §2 · p. 2
2.1 Problem Definition
- Given a task query and product-title corpus, the system generates catalog-oriented queries and BM25 rankings while seeking to preserve intent and narrow the lexical gap.Each reformulation produces a corresponding product ranking list.Source: §2.1 · p. 2
2.2 Catalog Semantic Anchoring
- Two catalog phrase resources support semantic-anchor construction and constrained reformulation using phrases drawn from product titles and broader corpus fields.Title-derived phrases are normalized into product categories, objects, or accessories, while non-product phrases such as brands, scenes, and marketing text are filtered.Source: §2.2 · pp. 2–3
2.3 Unified Five-Candidate Rewriting
- Five candidates preserve the original query, use BM25 feedback or role-based judgments, and offer conservative need-aware or compact catalog-phrase rewrites.The candidates represent distinct reformulation hypotheses that are evaluated jointly.Source: §2.3 · pp. 3–4
2.4 Protocol-Aware Selection
- Candidates are scored using BM25 results and reranking; Automatic selects the highest-quality candidate, while Interactive balances set quality against result overlap.BM25 supplies the top 100 products, reranking scores the top 10, and the proxy uses up to five discounted scores.Source: §2.4 · pp. 4–5
2.5 Final Retrieval Output
- After selection, BM25 runs independently for each reformulation, and the corresponding product-ranking lists remain separate for evaluation.Automatic returns one ranking list per original query; Interactive returns up to four.Source: §2.5 · p. 5
3 Experiments
TaskBridge is evaluated on the SIGIR 2026 eCommerce Query Reformulation Task under Automatic and Interactive protocols, using catalog resources, five-candidate rewriting, and BM25 retrieval. The reported results show improvements over the original-query baseline in the Open Phase and higher Interactive than Automatic scores in the Closed Phase.
3.1 Challenge Setup
- The task evaluates task-oriented e-commerce reformulation under Automatic, which permits one query, and Interactive, which permits up to four.TaskBridge uses the same product collection, queries, catalog resources, rewriting module, and BM25 retriever in both protocols; the retained query count differs.Source: §3.1 · p. 5
3.2 Implementation Details
- TaskBridge is an inference-time system without training or fine-tuning, using retrieval feedback and separate selection and execution stages.It uses up to two feedback rounds, retains the top 50 semantic-anchor candidates, reranks the top 10 products, and executes selected queries independently with BM25.Source: §3.2 · p. 5
3.3 Evaluation Metrics
- Evaluation reports Task Completion NDCG@10, Essential Product Recall@100, and MAP.These measure top-ten support for task completion, essential-product coverage among the top 100, and average precision over the ranking list, respectively.Source: §3.3 · pp. 5–6
3.4 Performance Comparison
- Open Phase TaskBridge scores are 0.49, 0.29, and 0.20 under Automatic and 0.58, 0.33, and 0.23 under Interactive, improving task-completion quality and overall ranking over the original-query BM25 baseline.The values correspond, in order, to Task Completion NDCG@10, Recall@100, and MAP.Source: §3.3 · p. 5, Table 1 · p. 6, §3.4 · p. 6
- In the Closed Phase, TaskBridge scores 0.27, 0.16, and 0.11 under Automatic, versus 0.31, 0.18, and 0.12 under Interactive.Interactive exceeds Automatic by 0.04, 0.02, and 0.01, respectively; the phases use different test sets and are not directly comparable.Source: Table 2 · p. 6, §3.4 · p. 6
4 Conclusion
TaskBridge combines catalog semantic anchoring, five complementary rewriting strategies, and protocol-aware selection to map implicit product needs to catalog expressions for BM25 retrieval. Interactive achieves higher scores than Automatic in both evaluation phases, suggesting complementary reformulations can better support task-oriented product retrieval.
4 Conclusion
- Interactive achieves higher scores than Automatic in both evaluation phases, suggesting complementary reformulations can better support task-oriented product retrieval.TaskBridge combines catalog semantic anchoring, five complementary rewriting strategies, and protocol-aware selection for BM25 retrieval.Source: §4 · p. 6