Combining Pretrained High-Resource Embeddings and Subword Representations for Low-Resource Languages