Event-enhanced Retrieval in Real-time Search

TL;DR

This work addresses semantic drift in embedding-based retrieval for real-time event search by introducing EER, a dual-encoder model augmented with a training-time generative decoder that extracts event triplets from document titles via prompt-guided generation. The model employs hard negative sampling, supervised contrastive learning, and pairwise losses, plus a query–generated-event relevance objective, to produce event-centric representations without increasing inference latency. Key contributions include a robust event-extraction decoder, prompt-guided generation, and a dedicated real-time event dataset with thorough ablations demonstrating performance gains over BM25, Sentence-BERT, and BGE baselines. Experimental results show significant improvements in Recall@10, MRR@10, and AUC, validating EER’s potential to improve real-time, event-focused information retrieval in practical search systems.

Abstract

The embedding-based retrieval (EBR) approach is widely used in mainstream search engine retrieval systems and is crucial in recent retrieval-augmented methods for eliminating LLM illusions. However, existing EBR models often face the "semantic drift" problem and insufficient focus on key information, leading to a low adoption rate of retrieval results in subsequent steps. This issue is especially noticeable in real-time search scenarios, where the various expressions of popular events on the Internet make real-time retrieval heavily reliant on crucial event information. To tackle this problem, this paper proposes a novel approach called EER, which enhances real-time retrieval performance by improving the dual-encoder model of traditional EBR. We incorporate contrastive learning to accompany pairwise learning for encoder optimization. Furthermore, to strengthen the focus on critical event information in events, we include a decoder module after the document encoder, introduce a generative event triplet extraction scheme based on prompt-tuning, and correlate the events with query encoder optimization through comparative learning. This decoder module can be removed during inference. Extensive experiments demonstrate that EER can significantly improve the real-time search retrieval performance. We believe that this approach will provide new perspectives in the field of information retrieval. The codes and dataset are available at https://github.com/open-event-hub/Event-enhanced_Retrieval .

Figures (5)

• Figure 1: An event corresponds to various queries and documents. Most queries are always concise, focusing on the key information of the event, and often contain abbreviations, omissions, grammatical irregularities, etc. For example, in the second query "mate60pro", "Huawei" is omitted, and "Mate" is entered as "mate". The document title is lengthy, contains redundant information, and the expression style is diversified. In the third title, the action "稳了 (Steady)" lacks a subject and is an unconventional syntax. The fourth title contains a lot of tags with "#". It is therefore difficult to relate queries to documents. The data here is from the real world.
• Figure 2: Architecture of the proposed EER model.
• Figure 3: The t-SNE visualization of representations from encoders without and with contrastive learning. As demonstrated in the left part, without contrastive learning, the model encodes queries into a smaller space with more collapses. And on the right, the addition of contrastive learning expands the embedding space with better alignment and uniformity.
• Figure 4: Typical case demonstration of EER and baseline. Relevant query-title pairs are marked as 1 and irrelevant ones as 0.
• Figure 5: Distribution of Roberta (a) and our method (b).