1 Predicting Actions with Commonsense Knowledge: an Application in Event-Driven Ad Placement Hogun Park, Heung-Seon Oh, Jihee Ryu, Sung-Hyon Myaeng School of Engineering, Information and Communications University 119, Munjiro, Yuseong-gu, Daejeon, 305-732, South Korea {gsgphg, ohs, zzihee5, myaeng} @ icu.ac.kr ABSTRACT Inferring or predicting actions to be taken by the user of a system is essential to the development of intelligent user interfaces. We employ ConceptNet, a large-scale commonsense knowledge base, and its derivative called EventNet to predict events associated with a given event. To validate the usefulness of inferred events, we developed a contextual ad placement system that accepts news articles and retrieves relevant ads. By predicting what actions the interested readers would take and what objects or entities are related to the actions, we attempt to get to the needs and intents of potential beneficiaries of ads, moving one step further beyond topical similarities used in modern ad placement systems. Related objects to be used for ad searching are found by using ConceptNet and WordNet. Preliminary experiments show that the proposed approach based on events and predicted actions retrieve unique sets of relevant ads that are not retrieved at all by either keyword-based or taxonomy-based semantic retrieval methods, validating our hypothesis that implicit user needs can be analyzed by inferring the actions to be taken. Author Keywords Action Prediction, Commonsense Knowledge, ACM Classification Keywords H5.2. [Information interfaces and presentation]: User Interface. – Graphical user interfaces. INTRODUCTION Inferring or predicting actions to be taken by the user of a system is essential to the development of intelligent user interfaces. For generlizability and domain independence, we employ ConceptNet [7], a large-scale commonsense knowledge base, and its derivative called EventNet [10] to predict actions associated with a given event. This capability of inferring the next or related actions and associated objects have a great potential for understanding user intent and the next actions expected of from the underlying system. To validate the usefulness of predicted actions and associated objects, we developed a contextual ad placement system that accepts news articles and retrieves relevant ads to be placed in them. Contextual advertising (or context match) recommends ads relevant to a web page. In many user studies [5][6], it has been shown that the contextual advertising optimizes the probability of clicks of Ads, so most commercial intermediaries, which play a role of a mediator between the advertiser and the publishers, have employed the contextual advertising. For placing more contextually relevant ads, many approaches were attempted to estimate the probability of clicks of ads to be displayed in a generic web page. The approaches presented in previous work are categorized into the two types: 1. Keyword-based ads recommendation places ads based on co-occurrence of keywords between a web page and a description or bid phrases of an ad. Term vectors are constructed for them and matched to estimate their similarity. Keyword-based ads recommendation methods suffer from a problem in semantic disambiguation, which needs keyword expansion for reducing a vocabulary impedance problem [3]. 2. Taxonomy-based ad recommendation is to solve the limitation of keyword-based ad placement. A recent work by Broder et al. [2] introduced semantic matching using a commercial taxonomy in addition to an existing keyword-based approach, but it is still limited because their matching is based on only topical features. It would be impossible, for example, to place an ad for a flower shop to a Web page announcing graduation ceremony unless some words semantically related to flowers are found in the page. To alleviate the problems associated with the two approaches, we propose an ad placement system that infers actions to be taken by the readers of the page and retrieve ads based on them. By predicting what actions the interested readers would take and what objects or entities are related to the actions, we attempt to get to the needs and intents of potential beneficiaries of ads, moving one step