An algorithm for cross-lingual sense-clustering tested in a MT evaluation setting

Unsupervised sense induction methods offer a solution to the problem of scarcity of semantic resources. These methods automatically extract semantic information from textual data and create resources adapted to specific applications and domains of interest. In this paper, we present a clustering algorithm for cross-lingual sense induction which generates bilingual semantic inventories from parallel corpora. We describe the clustering procedure and the obtained resources. We then proceed to a large-scale evaluation by integrating the resources into a Machine Translation (MT) metric (METEOR). We show that the use of the data-driven sense-cluster inventories leads to better correlation with human judgments of translation quality, compared to precision-based metrics, and to improvements similar to those obtained when a handcrafted semantic resource is used