Bing’s dynamic approach to whole page relevance

In an effort to enhance its search results quality, Bing is bringing about an array of changes, encompassing not just webpages, but also including news items, images, maps, videos and other relevant media objects in answers thrown to specific user queries. As it readies to take on Google and other engines, Bing is infusing more zing into those ubiquitous blue links along with summaries.

Richer experience for search engine users

Apart from look and feel, the emphasis is now on more depth and insight in keeping with the continued evolution of the Web. Their research team is looking to create a much richer overall experience for users. For instance, some results will consume more screen real estate, making it convenient for them to view the content. The topmost obvious answer for generic query is larger with vertically-arrayed deep links, other quick access links following it. This pattern is based on an assumption that for a common query, people are interested in the first answer itself. There is a news answer highlighted after the first webpage-based result.

Enhancing search results page attributes

The richness and intuitiveness of the search results page leads to many novel and interesting, search quality queries: Under what circumstances are you better served by an answer as against a traditional link to a webpage? How is the process rationalized in order to make a result more prominent? Bing researchers term such optimization problems ‘Whole Page Relevance’. They have developed a few methodologies to deal with some of the broader issues. A case in point is the Bing technology to blend the blocks of content – webpages and answers – into one single result set, called ‘Answer Ranking’.

To enhance the quality and relevance, Bing considers the search engine user behavior pattern in the real world. The way they respond to changes, the engineers assume a better blending algorithm will push users’ clicks towards the top of the results page. Such algorithm will promote a relevant answer on the specific result page upward as long as its win rate remains above 0.5. Online experiments are done based on this metric and subsequent results of competing blending algorithms are compared, to generate a realistic data set.

To improve this metric not only on basis of historical click data, but also for rare/ unusual queries, or to those with no previous history, other kinds of inputs are added, namely:

1. Confidence scores from the answer provider

2. Query characterizations

3. Features extracted from other answers and web pages, which will be shown on the page.

Elaborating on the changes, Bings’ Chief Scientist for Core Search, Dr. Jan Pedersen states, “Recently we have focused on new inputs, which can improve our ability to place temporally relevant answers. For instance, the news-based answer is designed for timeliness and speed. The aim is for it to update as fast as possible once a news story breaks and then declines in prominence gradually. Our methodology has succeeded in improving overall whole page relevance even while being well adaptable to new, different sorts of answers. Currently we deploy and maintain dozens of machine learned blending functions specialized to different answer types.”

The end effect of all this is users can get a rich set of search results, which are statistically more likely to be closer to exactly what they are looking for – whether it’s a text link, a news update, a map, a video, or snippet of information.