Molecular representation learning with language models and domain-relevant auxiliary tasks