English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Search query classification, as an effective way to understand user intents,is of great importance in real-world online ads systems. To ensure a lowerlatency, a shallow model (e.g. FastText) is widely used for efficient onlineinference. However, the representation ability of the FastText model isinsufficient, resulting in poor classification performance, especially on somelow-frequency queries and tailed categories. Using a deeper and more complexmodel (e.g. BERT) is an effective solution, but it will cause a higher onlineinference latency and more expensive computing costs. Thus, how to juggle bothinference efficiency and classification performance is obviously of greatpractical importance. To overcome this challenge, in this paper, we proposeknowledge condensation (KC), a simple yet effective knowledge distillationframework to boost the classification performance of the online FastText modelunder strict low latency constraints. Specifically, we propose to train anoffline BERT model to retrieve more potentially relevant data. Benefiting fromits powerful semantic representation, more relevant labels not exposed in thehistorical data will be added into the training set for better FastText modeltraining. Moreover, a novel distribution-diverse multi-expert learning strategyis proposed to further improve the mining ability of relevant data. By trainingmultiple BERT models from different data distributions, it can respectivelyperform better at high, middle, and low-frequency search queries. The modelensemble from multi-distribution makes its retrieval ability more powerful. Wehave deployed two versions of this framework in JD search, and both offlineexperiments and online A/B testing from multiple datasets have validated theeffectiveness of the proposed approach.

Towards Better Query Classification with Multi-Expert Knowledge Condensation in JD Ads Search