Using semantic data to improve cross-lingual linking of article clusters

This paper presents a system that uses semantic data to improve cross-lingual linking of news article clusters. Two approaches are compared. The first based on two different Canonical Correlation Analysis (CCA) feature vector definitions: MAX-CCA and SUM-CCA, whereas the second one has been developed using a better-performed CCA approach in combination with Entity vectors. The aim of the comparison was to determine whether taking into account the semantic aspect of news increases performance and improves linking. Evaluations of the aforementioned techniques on a news corpus, both against Google News and manual, revealed good performance of our system. The overall gain in precision and recall when using entity vectors was significant.