LexLattice
A multilingual extractive summarization model for legal texts based on neural cellular automata. It models the hierarchical structure of legal documents as a two-dimensional semantic lattice and runs masked 2D neural cellular automata over it for semantic integration, enabling efficient, high-quality legal text summarization.
Timeline (last 90 days)
LexLattice concretizes the hierarchical structure of legal documents into a two-dimensional semantic lattice and performs integration over the lattice before selecting summary sentences.
LexLattice uses a masked two-dimensional neural cellular automaton as the integration tool, allowing semantic information to repeatedly propagate and aggregate among adjacent cells in the lattice.
LexLattice achieves state-of-the-art ROUGE scores on all 24 languages of the EUR-Lex-Sum dataset under both multilingual and cross-lingual settings.
LexLattice has only 1.8 million total trainable parameters, concentrated in a single integrator, with its underlying multilingual encoder frozen.
LexLattice outperforms instruction-tuned baseline models with billions of parameters.
After training the integrator only on high-resource languages and transferring it to languages unseen during training, the researchers achieved a retention rate as high as 0.99.
The paper argues that LexLattice's integrator operates on language-agnostic semantic geometry rather than on the surface forms of specific languages, and therefore possesses cross-lingual generalization capability.
The paper argues that explicit structural modeling can replace indiscriminate scaling, achieving better and more traceable results at minimal computational cost.
All Facts (8)
LexLattice concretizes the hierarchical structure of legal documents into a two-dimensional semantic lattice and performs integration over the lattice before selecting summary sentences.
50%UnverifiedLexLattice uses a masked two-dimensional neural cellular automaton as the integration tool, allowing semantic information to repeatedly propagate and aggregate among adjacent cells in the lattice.
50%UnverifiedLexLattice achieves state-of-the-art ROUGE scores on all 24 languages of the EUR-Lex-Sum dataset under both multilingual and cross-lingual settings.
50%UnverifiedLexLattice has only 1.8 million total trainable parameters, concentrated in a single integrator, with its underlying multilingual encoder frozen.
50%UnverifiedLexLattice outperforms instruction-tuned baseline models with billions of parameters.
50%UnverifiedAfter training the integrator only on high-resource languages and transferring it to languages unseen during training, the researchers achieved a retention rate as high as 0.99.
50%UnverifiedThe paper argues that LexLattice's integrator operates on language-agnostic semantic geometry rather than on the surface forms of specific languages, and therefore possesses cross-lingual generalization capability.
50%UnverifiedThe paper argues that explicit structural modeling can replace indiscriminate scaling, achieving better and more traceable results at minimal computational cost.
50%