ExpandNet: A Dictionary-Based Cross-Lingual Word Sense Projection System

The ExpandNet system, open-sourced by the NLP Lab at the University of Alberta, aims to address the scarcity of semantic annotation resources in the field of multilingual NLP. Through a three-step translation-alignment-projection process, the system automatically converts source language vocabulary and semantic annotations into equivalent forms in the target language, supporting the expansion of multilingual semantic resources. Related work was included in the proceedings of the 39th Canadian Artificial Intelligence Conference (Canadian AI 2026).

In the field of natural language processing, the scarcity of semantic annotation resources (such as word sense disambiguation corpora) restricts the development of multilingual NLP. Most high-quality annotated datasets are concentrated in a few languages like English, while thousands of languages lack corresponding resources. How to transfer semantic knowledge from resource-rich languages to resource-poor ones is one of the core challenges in computational linguistics, and the ExpandNet system was designed to address this.

ExpandNet adopts a three-step process:

1. Sentence Translation: Supports Helsinki-NLP neural network models and OpenAI GPT series; manual translation can be provided to skip the automatic step.
2. Word Alignment: A core component, provides SimAlign and DBAlign algorithms; DBAlign requires bilingual dictionary guidance to improve accuracy.
3. Semantic Projection: Uses alignment information to project semantic annotations; ensures quality through part-of-speech (POS), named entity, dictionary filtering, and out-of-vocabulary (OOV) word processing.

ExpandNet's features include:

• Modular Design: The three-step process can run independently, supporting optimization or replacement of specific steps.
• Multilingual Support: Built-in spaCy models for English, Spanish, French, Chinese, and other languages.
• Evaluation Toolchain: Generates target language gold standard data via BabelNet, supporting quantitative evaluation of projection quality.

ExpandNet application scenarios:

• Academic Research: Provides standardized tools and benchmarks for cross-lingual semantic research.
• Industrial Applications: Reduces the cost of multilingual NLP systems, eliminating the need to annotate training data from scratch.
• Specific Scenarios: Multilingual word sense disambiguation, cross-lingual retrieval, machine translation improvement, and low-resource language processing.

ExpandNet is implemented in Python, relying on spaCy to complete basic NLP tasks and supporting multi-process parallelism to improve efficiency. For multi-word expressions, tokenized words with spaces are connected using underscores; the projection step adopts a conservative strategy to prioritize the credibility of results. Each step provides detailed command-line parameters, allowing users to adjust flexibly.

ExpandNet is an important advancement in the field of cross-lingual semantic transfer, balancing accuracy and flexibility through dictionary-driven alignment and projection methods. Its modular design and comprehensive documentation provide a foundation for subsequent improvements. In the future, with the accumulation of more language dictionary resources and the improvement of translation quality, it is expected to play a greater role in building multilingual semantic infrastructure.