GPU-accelerated lemmatization with CUDA and RAPIDS accelerator for Spark.
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Lemmatization is one of the few preprocessing stages in NLP pipelines that is still predominantly CPU-bound, even as tokenization, vectorization, and model inference increasingly rely on high-throughput GPU execution. Deterministic, dictionary-based morphological analyzers deliver stable and linguistically correct results, but their sequential nature, complex branch divergence, and interoperability overhead significantly limit processing throughput in large-scale distributed systems. In this work, we propose a GPU-native deterministic lemmatizer for Ukrainian, implemented as a custom CUDA kernel and integrated into Apache Spark via a RAPIDS-accelerated JNI UDF. A large morphological dictionary (VESUM/dict_uk) is compiled into a compact trie layout optimized for coalesced memory traversal and fully parallel token processing, preserving accuracy while eliminating CPU constraints and reducing end-to-end latency. We evaluate three Spark pipelines — Python with pymorphy3, Java with Morfologik, and our CUDA implementation — on corpora up to 175.6M tokens. The GPU approach achieves 35.77M tokens/s, outperforming the optimized Java baseline by 8.8× and the Python baseline by up to 120×. Our results demonstrate that classical morphological analysis — historically treated as inherently sequential — can be re-architected into an efficient, scalable, and bandwidth-bound GPU operation suitable for real-world big-data workflows.
Fields of Study & Subjects
Parallel Computing and Optimization Techniques
Natural Language Processing Techniques
Topic Modeling
Publication details
- Journal / source
- COMPUTER-INTEGRATED TECHNOLOGIES EDUCATION SCIENCE PRODUCTION
- Issue
- 62
- Pages
- 89–98
- Publication date
- 28 March 2026
- Language
- EN