Edgebert: a Click-Graph Enhanced Approach for Search Ads Matching

Sponsored search advertisements appear alongside search results when users search for the services and products listed on search engines. And these ads have become one of the most profitable marketing channels. Relevance matching, the core concept of the Search Engine Advertising, has garnered increasing attention owing to its considerable research challenges and substantial practical significance. Given the brevity of both queries and ads, as well as the limited semantic information available, a large volume of manually labeled data is essential for training an effective relevance model. However, manually collecting sufficient annotations is both time-consuming and costly. Therefore, our approach leverages the extensive click data from search logs to complement the relevance model and improve its performance. The query-ad pair is represented by two sub-click-graphs, each centered around the query and the ad, respectively. These sub-graphs are expected to provide richer semantic information compared to the original input, which consists of a single sentence. In this paper, we propose two graph-based models, TokenNet and EdgeBert, which explore click data from the perspectives of tokens and segments, respectively. And we evaluate the EdgeBert model on a real-world dataset. The experimental results demonstrate the superiority of our approach.