It is necessary to understand the content of articles and user preferences to make effective news recommendations. While ID-based methods, such as collaborative filtering and low-rank factorization, are well known for making recommendations, they are not suitable for news recommendations because candidate articles expire quickly and are replaced with new ones within short spans of time. Word-based methods, which are often used in information retrieval settings, are good candidates in terms of system performance but have issues such as their ability to cope with synonyms and orthographical variants and define "queries" from users' historical activities. This paper proposes an embedding-based method to use distributed representations in a three step end-to-end manner: (i) start with distributed representations of articles based on a variant of a denoising autoencoder, (ii) generate user representations by using a recurrent neural network (RNN) with browsing histories as input sequences, and (iii) match and list articles for users based on inner-product operations by taking system performance into consideration. The proposed method performed well in an experimental offline evaluation using past access data on Yahoo! JAPAN's homepage. We implemented it on our actual news distribution system based on these experimental results and compared its online performance with a method that was conventionally incorporated into the system. As a result, the click-through rate (CTR) improved by 23% and the total duration improved by 10%, compared with the conventionally incorporated method. Services that incorporated the method we propose are already open to all users and provide recommendations to over ten million individual users per day who make billions of accesses per month.

Contextual advertising is a form of textual advertising usually displayed on third party Web pages. One of the main problems with contextual advertising is determining how to select ads that are relevant to the page content and/or the user information in order to achieve both effective advertising and a positive user experience. Typically, the relevance of an ad to page content is indicated by a tf-idf score that measures the word overlap between the page and the ad content, so this problem is transformed into a similarity search in a vector space. However, such an approach is not useful if the vocabulary used on the page is expected to be different from that in the ad. There have been studies proposing the use of semantic categories or hidden classes to overcome this problem. With these approaches it is necessary to expand the ad retrieval system or build new index to handle the categories or classes, and it is not always easy to maintain the number of categories and classes required for business needs. In this work, we propose a translation method that learns the mapping of the contextual information to the textual features of ads by using past click data. The contextual information includes the user's demographic information and behavioral information as well as page content information. The proposed method is able to retrieve more preferable ads while maintaining the sparsity of the inverted index and the performance of the ad retrieval system. In addition, it is easy to implement and there is no need to modify an existing ad retrieval system. We evaluated this approach offline on a data set based on logs from an ad network. Our method achieved better results than existing methods. We also applied our approach with a real ad serving system and compared the online performance using A/B testing. Our approach achieved an improvement over the existing production system.

In this paper, we report our system that disambiguates person names in Web search results. The system uses named entities, compound key words, and URLs as features for document similarity calculation, which typically show high precision but low recall clustering results. We propose to use a two-stage clustering algorithm by bootstrapping to improve the low recall values, in which clustering results of the first stage are used to extract features used in the second stage clustering. Experimental results revealed that our algorithm yields better score than the best systems at the latest WePS workshop.