LexLattice: Multilingual Extractive Summarization via Neural Cellular Automata on Document Hierarchies

Summary:ArXiv introduces LexLattice, an innovative multilingual extractive summarization model for legal texts. By representing the hierarchy of legal acts as a two-dimensional semantic lattice and consolidating information using a masked 2D neural cellular automata, LexLattice achieves state-of-the-art performance across 24 languages in both multilingual and cross-lingual settings. Despite having only 1.8M trainable parameters, it surpasses models with billions of parameters, demonstrating its efficien

---

Background and Challenges

In the field of legal text processing, the faithfulness of summarization is crucial. Traditional methods typically employ extractive summarization by ranking paragraphs or other structural units in isolation, which often fails to effectively integrate key information dispersed across different parts of the text, thus affecting the completeness and accuracy of the summary.

Core Innovations of LexLattice

LexLattice addresses these challenges through the following innovations:

1. Semantic Lattice Representation: The model transforms the hierarchy of legal texts into a two-dimensional semantic lattice, allowing it to capture semantic relationships and hierarchical information within the text.
2. Neural Cellular Automata Consolidation: Utilizing a masked 2D neural cellular automata for information consolidation ensures that the model can effectively integrate cross-paragraph key information when selecting summary content.
3. Cross-Lingual Transfer Capability: The model demonstrates strong cross-lingual transfer capabilities, achieving near-lossless performance retention even when trained only on high-resource languages.

Experimental Results and Performance

LexLattice was tested on the EUR-Lex-Sum dataset across 24 languages, with the following results:

• In both multilingual and cross-lingual settings, LexLattice's ROUGE scores are the highest, surpassing models with billions of parameters.
• The model exhibits near-lossless performance retention even in low-resource languages, validating its cross-lingual transfer capabilities based on semantic geometry rather than surface form.

Technical Highlights

• Efficient Consolidation: Through the semantic lattice and neural cellular automata, LexLattice can efficiently consolidate cross-paragraph key information, enhancing the completeness and accuracy of the summary.
• Cross-Lingual Adaptation: The model's strong cross-lingual adaptation capabilities make it an ideal choice for processing multilingual legal texts.
• Lightweight Model: With only 1.8M parameters, yet outperforming models with billions of parameters, LexLattice demonstrates its efficiency and scalability.

Industry Impact and Developer Recommendations

The release of LexLattice brings a new technological breakthrough to the field of legal text processing, with broad application prospects in multilingual environments. For developers, LexLattice can be applied in the following scenarios:

• Multilingual Legal Document Processing: Utilize LexLattice's efficient summarization capabilities to improve the processing efficiency of multilingual legal documents.
• Cross-Lingual Information Retrieval: Combine LexLattice's cross-lingual transfer capabilities to achieve accurate and efficient cross-lingual information retrieval.

Conclusion

LexLattice, through its innovative semantic lattice and neural cellular automata technologies, demonstrates strong performance in the field of multilingual legal text summarization, providing a new direction for future research and technological applications.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-09-24)

— END —

Tags: #Multilingual Model #Legal Text Processing #Neural Cellular Automata #Cross-Lingual Transfer #Semantic Lattice