Abstract
Entity recommendation, providing search users with an improved experience via assisting them in finding related entities for a given query, has become an indispensable feature of today's search engines. Existing studies typically only consider the queries with explicit entities. They usually fail to handle complex queries that without entities, such as "what food is good for cold weather", because their models could not infer the underlying meaning of the input text. In this work, we believe that contexts convey valuable evidence that could facilitate the semantic modeling of queries, and take them into consideration for entity recommendation. In order to better model the semantics of queries and entities, we learn the representation of queries and entities jointly with attentive deep neural networks. We evaluate our approach using large-scale, real-world search logs from a widely used commercial Chinese search engine. Our system has been deployed in ShenMa Search Engine and you can fetch it in UC Browser of Alibaba. Results from online A/B test suggest that the impression efficiency of click-through rate increased by 5.1% and page view increased by 5.5%.
Introduction
Traditional entity recommendation systems are primarily "entity-centric." When a user searches for a specific entity like "Inception," suggesting related entities like "Christopher Nolan" is straightforward. However, Alibaba's internal data reveals that over 50% of user queries are "complex"—they describe a need or context without naming a specific entity (e.g., "how to lose weight fast").
Existing models often struggle with these queries because they lack a semantic bridge to connect descriptive natural language to specific entities in a Knowledge Graph. This research addresses the "implicit user need" problem by framing entity recommendation as a large-scale semantic matching task rather than a simple keyword lookup.
Methodology: Deep Collaborative Matching
The core of the proposed system is a Deep Collaborative Matching framework. While initial iterations used basic word embeddings and n-gram features, the final production model utilizes an Enhanced Deep Match Model (att-BiLSTM).
att-BiLSTM Architecture
To capture the global temporal and structural information of a query, the model employs a Bi-directional LSTM (BiLSTM). This allows the system to understand word order and context more effectively than simple bag-of-words models. Furthermore, a Self-Attention Mechanism is integrated to automatically weigh the importance of different words. For instance, in the query "what food is good for cold weather," the attention layer prioritizes "food" and "cold" while de-emphasizing functional words.
Conceptualized Recommendation
To enhance the user interface, the system leverages a "Cognitive Concept Graph." This allows the engine to cluster recommended entities into logical categories. Instead of a flat list, users see entities grouped under concepts (e.g., "Food"), making the recommendations more interpretable and useful. The system is optimized for extreme efficiency, generating query embeddings in under 5ms to meet real-time search demands.
Evaluation and Impact
The model was trained on billions of query-entity pairs derived from user clicks, document metadata, and manual tags from the ShenMa Search Engine.
Offline Results
The att-BiLSTM model demonstrated superior performance over standard DNN and n-gram baselines. It achieved a Precision@10 of 30.95, a significant improvement over the base DNN's 28.29, proving the value of global context and attention mechanisms.
Online Deployment
The system was deployed on Alibaba’s ShenMa Search Engine within the UC Browser. In a 7-day online A/B test, the context-aware model led to a 5.1% increase in Click-Through Rate (CTR) and a 5.5% increase in Page Views (PV). These results confirm that better semantic understanding of complex queries directly translates to higher user engagement.
Conclusion
By moving beyond keyword matching to joint semantic representation, Alibaba's context-aware model successfully addresses the challenge of implicit user needs in search. Future work will explore more advanced encoders like BERT and the integration of multi-modal features to further refine the recommendation experience.
Summary updated