Personalized NewsEvent Retrieval for Small Talk in Social Dialog Systems

This paper presents the NewsTeller system which retrieves a news event based on a user query and the user’s general interests. It can be used by a social dialog system to initiate news-related small talk. The NewsTeller system is implemented as a pipeline with four stages: After collecting a large set of potentially relevant news events, a classifier is used to filter out mal- formed events. The remaining events are then ranked ac- cording to a relevance value predicted by a regressor. In a final step, a short summary of the highest-ranked event is generated and returned to the user. Both the classifier and the regressor were evaluated on hand-labeled data sets. In addition to this, a user study was conducted to further validate the system. Evaluation results indicate that the proposed approach performs significantly better than a random baseline.