Selected publications

Abstract summary: Introduces the Multilingual Semantic Layer and a semi-automatic multilingual dataset creation pipeline, then shows that manually refining 1,100 sentences across 11 languages helps reduce the semantic parsing performance gap for low-resource languages.

Abstract summary: Presents a large multilingual corpus with hundreds of millions of silver semantic annotations across four NLU tasks and five languages, aiming to make explicit semantic supervision more available across languages and tasks.

Abstract summary: Proposes a compact AMR linearization and a modular parser that cuts token counts by roughly 40 to 50 percent and reduces training and inference time by about 80 percent while preserving strong parsing quality.

Abstract summary: Builds a cross-lingual AMR aligner that recovers span-to-graph alignments directly from transformer cross-attention, avoiding English-specific rules and improving alignment quality across multiple languages.

Abstract summary: Examines the weaknesses of existing AMR ensemble methods, adds validation for structural constraints, and proposes two transformer-based ensemble strategies that are both more robust and less computationally expensive.

Abstract summary: Introduces LeakDistill, which injects graph structure into transformer representations through structural adapters and word-to-node alignment, leading to state-of-the-art AMR parsing without extra training data.

Abstract summary: Introduces the BabelNet Meaning Representation as an interlingual formalism built on BabelNet and VerbAtlas, releases the BMR 1.0 dataset, and shows that the formalism supports strong multilingual parsing and generation.

Abstract summary: Presents BMR as a broad, fully semantic and language-independent meaning representation designed to connect text, images, audio, video, and logical forms across languages and AI settings.