Learning Cross-Lingual Sentence Representations via a Multi-task Dual-Encoder Model