This article “Creating a Hybrid Rule and Neural Network Based Semantic Tagger Using Silver Standard Data: The PyMUSAS Framework for Multilingual Semantic Annotation”, was developed by Andrew Moore, Paul Rayson, et al. and published in the proceedings of LREC 2026.
This paper presents the first large-scale, multilingual evaluation of semantic tagging under the USAS framework, comparing traditional rule-based taggers against new neural network models across five languages (English, Irish, Finnish, Welsh, and Chinese). To overcome the lack of manually labeled training data, the researchers used the existing rule-based English tagger to automatically generate a silver-labeled training set of over 5 million words, then used it to train several neural models, which they combined with the rule-based system into a hybrid tagger that falls back on the neural model whenever a word is missing from the dictionary. The results showed that neural and hybrid models consistently outperformed rule-based tagging alone across all five languages — even for languages the neural models were never explicitly trained on — demonstrating effective cross-lingual transfer from English-only training data. As part of the work, the team also released the first open, manually annotated Chinese dataset for USAS tagging, along with all trained models, training data, and code as public resources for the research community.
