Tailor knowledge graph for query understanding: linking intent topics by propagation

Shi Zhao, Yan ZhangEMNLP 2014 · DOI

Introduction

Search engine queries are often "laconic"—short, underspecified, and lacking context. While modern search systems leverage Knowledge Graphs (KGs) like Freebase to interpret entities (e.g., "Toyota"), they frequently overlook the surrounding "refiners" (e.g., "for sale") and the rich intent signals embedded in clicked URLs. This research introduces Tailor, a framework designed to bridge this gap by expanding traditional KGs to include non-entity words and user behavior as "intent topics."

The Tailor Framework

The core innovation of Tailor is the unification of entities, refiners, and URLs into a single graph structure. By treating these disparate elements as nodes in a common semantic space, the system can leverage graph-based algorithms to infer the meaning of previously "invisible" query components.

Graph Construction and Label Propagation

The process begins with the construction of a massive graph where edges represent co-occurrence within search sessions or semantic similarity. To manage the scale and noise of web data, Locality Sensitive Hashing (LSH) is employed to merge near-duplicate nodes.

The "tailoring" occurs through an unsupervised label propagation algorithm. Semantic features (topic and domain vectors) are initialized for known entities using Freebase. These features then "flow" through the graph to unlabeled refiner and URL nodes. For example, if the term "lyrics" frequently appears alongside music entities, the "lyrics" node eventually inherits "music" and "entertainment" attributes from its neighbors.

Experimental Setup

The authors evaluated Tailor using the AOL Search Logs (20 million queries) and Freebase (7 million facts). Initial entity recognition was handled by the Tagme tool. A key motivation for this work is the discrepancy between the distribution of topics in static knowledge bases versus actual user queries.

Figure 1: Distribution of topics across query sets and Freebase knowledge sources.

As shown in Figure 1, certain domains like "music" are overrepresented in Freebase facts compared to their frequency in actual user query sessions, highlighting the need for a mechanism that adapts global knowledge to local search contexts.

Results and Discussion

The evaluation focused on a session intent understanding task, comparing the proposed ECR (Entity + Clicks + Refiners) model against entity-only and popularity-based baselines.

The ECR model demonstrated significant improvements, achieving a MAP@5 of 0.722, compared to 0.676 for the entity-only approach. The use of Geometric Mean Average Precision (GMAP) further highlighted the model's robustness, showing that the inclusion of refiners and clicks provides a more stable representation of user intent across diverse sessions.

Figure 2: Performance sensitivity (MAP and GMAP) across varying weights for refiners and clicks.

Figure 2 illustrates the sensitivity of the model's performance to the weighting parameters for refiners and clicks. The heatmaps indicate that optimal performance is achieved when both signals are integrated, though the system remains relatively robust across a range of parameter values.

Conclusion and Future Work

Tailor successfully demonstrates that non-entity query components and user clicks can be semantically enriched through graph propagation. By "tailoring" a global KG with local session data, the system achieves a more nuanced understanding of search intent. Future research will explore automated methods for parameter weighting and deeper integration of granular KG properties to further refine intent modeling.

Summary updated

BibTeX
@inproceedings{zhao2014tailor,
  author = {Shi Zhao and Yan Zhang},
  title = {{Tailor knowledge graph for query understanding: linking intent topics by propagation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
  doi = {10.3115/v1/d14-1114},
}
Figure