This is the summary of an article by Hyo-Jung Oh, Changki Lee, and Chung-Hee Lee. You can get the pdf of the behavioral targeting article here: Analysis of the Empirical Effects of Contextual Match Advertising for Online News
There are two main types of online advertising: sponsored search and contextual advertising. Contextual advertising uses the rich content found in publisher pages and uses that to post relevant ads on the page. Sponsored ads need to rely more on short queries, e.g. Google AdWords.
Contextual advertising research has focused on means to extract keywords in pages, which faces lots of problems including the fact that some words or phrases do have various meanings. But this study introduces AdContX, a rich CM (contextual matching) model. This model is then used for online news in an attempt to study its empirical effects.
Practical Online Advertising Service Model
There are four main players in online contextual advertising: users, content provider, advertiser, and 0ad broker. The ad broker mediates the publisher and advertiser, and is responsible for selecting the ads that will be displayed on web pages. Advertisers pay for clicks on their advertisements for the desired traffic to their websites, content publishers get advertisement revenue, and the ad broker gets a share from this revenue. This is the Ad Broker model.
In this study, another player is added, called the Contextual Matcher. This player coordinates ad matching based on content context.
AdContX: Contextual Matching Advertising
Ads located in a news document online are highly valuable if they pertain to products related to the context of the page. The following are are the primary components of AdContX: Linguistic Analysis, Classification Based on Ad-Taxonomy, Ad Query Generation, Semantic Ad Matching.
A web page published by a content publisher is sent to AdContX, which converts it into plain text. Part-of-speech tagging, named entity recognizing (NER), and morphological analysis is done to extract significant keywords. This study considers 147 named entity types arranged hierarchically. Names include time, date, civilization, location, organization and person. Named entities are very important in the accuracy of AdContX because most brands are proper names.
Classification Based on Ad Taxonomy
A taxonomy consisting of 35 coarse categories (e.g electronics, furniture, cars) and 435 fine classes was used for AdContX after referring to taxonomies in online ad agencies and shopping malls in Korea. The ads, like the web pages, have an ad-taxonomy classification from the bid-phrases the advertiser used for the campaign. As such, web pages and ads can be semantically matched.
Ad Query Generation
Contextual Matching aims to look for ad space and determine the best ads for a certain web page. So the next step is to extract key bid-phrases from the web page content. The number of queries is also extended using a linguistic resource containing a hundred thousand bid phrases.
Semantic Ad Matching
The Ad Match is stronger for a page and an advertisement if they have the same taxonomy node classification and weaker if the nodes are farther apart.
AdContX’s performance was compared with the traditional advertising model by using click-through rate (CTR) evaluations. In the methodology, a user that select a news service is given an advertisement which is randomly selected between AdContX and the traditional model.
Results show that AdContX performs better than the traditional model in terms of CTR comparisons by an improvement of 14 percent. In particular, the area of finance news is where AdContX performs best.