Leveraging Translations for Lexical Semantics
Arnob Mallik · ERA: Education and Research Archive (University of Alberta) · 2021
We leverage multilingual translations from parallel corpora to improve sense annotations, build end-to-end Word Sense Disambiguation pipelines and detect cross-lingual lexical entailment. Based on theories of translational equivalence, we propose novel algorithms capable of correcting noisy sense annotations on a parallel corpus. We show that, when applied to bilingual slices of a parallel corpus, these algorithms can rectify noisy sense annotations and thereby produce multilingual sense-annotated training data of improved quality. Furthermore, we propose novel end-to-end pipelines which can produce high-quality sense annotations from scratch in a fully unsupervised manner. Our methods achieve state-of-the-art results on standard WSD datasets for unsupervised approaches in several languages. Additionally, by exploring the generalization property of translations, we develop novel approaches to detect cross-lingual lexical entailment by leveraging word embeddings along with translations. We evaluate our methods on a standard shared task dataset and achieve encouraging results constituting a strong proof-of-concept. In summary, our results in three different tasks of lexical semantics confirm the utility of translations in this field.