Source-linked AI summary
VnCoreNLP: A Vietnamese Natural Language Processing Toolkit
Thanh Vu, Dat Quoc Nguyen, Dai Quoc Nguyen, Mark Dras, Mark Johnson
TL;DR
Vietnamese NLP lacked a unified pipeline for key benchmarked tasks, while an existing toolkit was too slow for large-scale use. VnCoreNLP addresses this gap with a fast Java toolkit covering core linguistic annotations and achieves higher results than previous published systems on the same benchmarks. Its word segmentation component reaches 97.90% F1 at 62K words per second.
Problem
Vietnamese NLP needed a unified pipeline for publicly benchmarked word segmentation, POS tagging, NER, and dependency parsing, while NNVLP processed only about 300 words per second.
Method
VnCoreNLP is a Java toolkit providing word segmentation, POS tagging, named entity recognition, and dependency parsing through existing feature-based and transformation rule-based components.
Results
VnCoreNLP components obtain higher results than previous published results on the same benchmark datasets; word segmentation reaches 97.90% F1 at 62K words per second.
Takeaways & Limitations
VnCoreNLP provides an easy-to-use, fast, and accurate Vietnamese NLP pipeline with rich linguistic annotations to facilitate Vietnamese NLP research.
Takeaways & Limitations
All reported processing speeds were computed on a personal computer with an Intel Core i7 2.2 GHz processor.
Abstract
from arXiv · showhide
We present an easy-to-use and fast toolkit, namely VnCoreNLP---a Java NLP annotation pipeline for Vietnamese. Our VnCoreNLP supports key natural language processing (NLP) tasks including word segmentation, part-of-speech (POS) tagging, named entity recognition (NER) and dependency parsing, and obtains state-of-the-art (SOTA) results for these tasks. We release VnCoreNLP to provide rich linguistic annotations to facilitate research work on Vietnamese NLP. Our VnCoreNLP is open-source and available at: https://github.com/vncorenlp/VnCoreNLP
1 Introduction
VnCoreNLP addresses the need for a Vietnamese NLP pipeline by combining key annotation tasks in an easy-to-use, fast, and accurate toolkit.
- Motivation: Vietnamese NLP has public benchmark datasets for word segmentation, POS tagging, dependency parsing, and NER, creating a need for an integrated pipeline.The proposed pipeline is intended to assist users, researchers, and downstream tool developers.
- Motivation: Earlier Vietnamese pipelines wrapped existing segmenters and taggers, but those components were no longer considered state of the art.
- Motivation: NNVLP applied a BiLSTM-CNN-CRF model but did not compare against state-of-the-art traditional models and processed only about 300 words per second.
- VnCoreNLP: VnCoreNLP provides word segmentation, POS tagging, NER, and dependency parsing through a Java NLP toolkit designed to facilitate Vietnamese NLP research.
- VnCoreNLP: VnCoreNLP components are packaged in one .jar file and can run from either the command line or the Java API.
- VnCoreNLP: VnCoreNLP is designed for large-scale data and users with limited computational resources.
- VnCoreNLP: VnCoreNLP components obtain higher results than all previously published results on the same benchmark datasets.
2 Basic usages
VnCoreNLP supports simple command-line and API-based annotation workflows, producing structured linguistic output for sentences and larger texts.
- Command-line usage: A minimal command runs VnCoreNLP on input.txt and writes annotations to output.txt.The command uses the VnCoreNLP JAR with a 2 GB heap setting.
- Command-line usage: Figure 2 presents the minimal command for running VnCoreNLP.
- Output: The example input sentence receives word segmentation, POS, NER, and dependency annotations in the output.The example is a Vietnamese sentence about working at Vietnam National University, Hanoi.
- Output: Table 1 represents each annotated token with word index, word form, POS tag, NER label, head index, and dependency relation type.
- Java API: The same annotation output can be obtained through the Java API by creating an Annotation object and calling pipeline.annotate.
- Java API: The API also supports processing paragraphs or entire news stories and retrieving individual sentences from an annotation.
3 Components
VnCoreNLP integrates existing high-performing models with a new word-segmentation model into one Vietnamese annotation pipeline.
- Design approach: The toolkit focuses on integrating existing models rather than developing a new approach for every component task.
- Design approach: Except for word segmentation, VnCoreNLP applies traditional feature-based models that achieved state-of-the-art results for English POS tagging, NER, and dependency parsing.The paper motivates feature-based models as fast and accurate for less-resourced languages such as Vietnamese.
- wseg: Word segmentation uses a transformation rule-based learning model because Vietnamese whitespace separates syllables as well as words.The model is reported to improve both segmentation accuracy and speed over previous word segmenters.
- pos: POS tagging uses MarMoT, a generic CRF framework and state-of-the-art POS and morphological tagger.
- ner: Named entity recognition uses a dynamic feature induction model that automatically optimizes feature combinations.
- parse: Dependency parsing uses a greedy transition-based parser with selectional branching.
4 Evaluation
VnCoreNLP is evaluated across Vietnamese word segmentation, POS tagging, NER, and dependency parsing, with comparisons covering traditional feature-based and neural models. The toolkit reports leading accuracy or F1 results together with high processing speeds.
- POS tagging: 95.88% accuracy and 25K words per second give VnCoreNLP the highest reported POS-tagging performance and fast processing.The component also outperforms BiLSTM-CRF-based models.
- Named entity recognition: VnCoreNLP’s NER evaluation compares SOTA feature-based and neural models, addressing a comparison not previously reported for Vietnamese NER.The compared neural baselines include BiLSTM-CRF variants using character representations or predicted POS tags.
- Named entity recognition: The NER evaluation replaces gold POS tags with automatically predicted tags and merges full-name syllables to create a more realistic pipeline setting.The dataset contains 16,861 training/development sentences and 2,831 test sentences, while the revised split uses 14,861 training and 2,000 development sentences.
- Named entity recognition: 88.55% F1 at 18K words per second gives VnCoreNLP the highest reported NER score while remaining substantially faster than the second most accurate model.The second most accurate model is BiLSTM-CRF + CNN-char; the evaluation uses the test set.
- Dependency parsing: VnCoreNLP achieves the highest reported dependency-parsing LAS and processes 8K words per second, including results without automatic NER features.The comparison uses LAS and UAS on the Vietnamese dependency treebank, with gold or automatically predicted POS tags.
5 Conclusion
VnCoreNLP is an easy-to-use, fast, and accurate Vietnamese NLP pipeline providing core linguistic processing steps. It is released for research and educational purposes, while future work will add Vietnamese linguistic features.
- VnCoreNLP provides word segmentation, POS tagging, NER, and dependency parsing in one Vietnamese NLP pipeline.
- The toolkit is easy to use, fast, and accurate, using existing predefined features in traditional feature-based models.
- Future work will incorporate Vietnamese linguistic features into the feature-based models.
- VnCoreNLP is released for research and educational purposes.