Explainable Projection: Cross-lingual semantic role labeling

Anonymous

Abstract: Semantic role labeling (SRL) is a central task in many applications, e.g., machine translation, question answering, summarization, and more recently, complex tasks such as stance detection. However, cross-lingual projection of SRL labels has remained a thorny problem in NLP. The scarcity of semantically annotated corpora makes it difficult to build semantic role labelers, particularly for languages where hand-annotated labels are not readily available. We leverage semantic isomorphism at the level of predicate-argument structure to induce SRL systems from unlabeled bilingual corpora. We demonstrate that this approach yields explainable representations that readily project to new languages. Our novel contribution is the use of a simple word-to-word alignment followed by a First Come First Assign (FCFA) algorithm and a handful of linguistically-informed constraints specified at the predicate-argument level. These constraints provide a systematic mapping to semantic-role divergence categories that serve as the basis for analysis of our FCFA approach. A two-step process rapidly produces explainable SRL output: simple alignment followed by application of FCFA. This approach yields SRL projection results that are comparable to state of the art performance (XSRL), but without relying on complex transformer-based scoring schemes for multi-word alignments.

Paper Type: long

Research Area: Multilinguality and Language Diversity