Bareun
Bareun is a Korean morphological analyzer engine and, at the same time, a Korean spell checker, built by bareun.ai.
It is written bareun in Latin letters. Readers who do not know Korean pronounce it bɑʀɯɴ or pɑʀɯɴ.
Morphological analysis is free to use
Bareun's morphological analysis is free for research use by individuals, companies, schools, and news organizations.
Bareun is built and provided jointly by the Bareun team (Baikal AI) and the Korea Press Foundation. It serves as the morphological analysis engine of BIGKINDS, the Korea Press Foundation's news big-data analysis system, and is open to news organizations, press-related public institutions, and also to individuals and groups doing research.
Commercial services and commercial use, however, require a paid license.
Commercial services
Commercial services include business services, and also any always-on service aimed at end users, even a non-commercial one. In other words, any continuous use of this product's features for purposes other than research is a commercial service.
Commercial research
Morphological analysis can also be used for research with a commercial purpose. One example is building large-scale, secondarily processed training data after the basic research is finished. In this case you must obtain a paid license before using it.
Spell checking
Bareun is also the name of a spell checker covering Korean grammar and usage, word spacing, and standard forms. Spell checking is a paid service. Even when you install it yourself, only users with a paid registration can use it.
A paid license is required in this case as well.
History of bareun.ai
Bareun started from the research of a linguist with deep knowledge of Korean, and grew into today's morphological analyzer and spell checker.
| Time | Milestone |
|---|---|
| 2018 | Prof. Jinho Park's paper published, the academic root of Bareun |
| 2019 / 2021 | Research and development at Baikal AI and Baikal NLP (1.4.2 and others) |
| 2023.2.23 | Bareun morphological analyzer 2.0 released, with continuous upgrades since |
| 2025 | Bareun spell checker 3.0.0 launched |
| Now | 3.4.0 released (2026-11-05) |
Analyzing Korean correctly
Bareun is built on a deep understanding of Korean. Rather than chasing new techniques, we went through the corpus line by line and refined it with care.
- Very high accuracy.
- Highly adaptable, handling a wide range of Korean expressions.
- Easy to use.
- Works for researchers and in enterprise environments alike.
Even in the current era dominated by large language models, Bareun is attempting a qualitative change across Korean natural language processing as a whole.
Transformer-based
The main reason Bareun stands apart is its Transformer technology. ChatGPT and BERT have shown striking results with Transformers. We hope Bareun will be a meaningful work among those that use Transformers.
- Bareun closely analyzes the characteristics of Korean inflection and conjugation and finds 115 segmentation rules. Of these, 14 are error-detection rules that let the spell checker catch morphological errors at the segmentation stage.
- In this process it identifies 8 major morphological units: substantive (N), particle (J), predicate (V), ending (E), adverbial (A), determiner (M), interjection (I), and symbol (S).
- Using these, we built a word embedding. It embeds
가#V,가#J, and가#Nseparately. - It finds morphemes according to the National Institute of Korean Language's 47 POS standard.
- It post-corrects the analysis with a custom dictionary that includes verbs and adjectives.
Transformers are used throughout this entire process.
What makes Bareun stand out
Key features
Morphological analysis
- Separates morphemes well.
- Also segments at the word-unit (eojeol) level.
- Adjusts eojeol boundaries by inserting or removing spaces.
- For irregular conjugation in particular, it reports the exact original position of each recovered form.
For example, when "제
킨데요." is analyzed as키+이+ㄴ데+요,키,이, andㄴ데all share the same start position.킨goes through a split into키+이+ㄴby Bareun's 104th segmentation rule,생략_모음체언_이_ㄴㄹㅂ_1. Bareun traces this process to output the exact position values. - Shows where each analyzed word came from. In particular, it distinguishes custom dictionary entries and neologisms.
Word segmentation
Outputs the result of the segmentation stage, before morphological analysis. It does not tag morphemes, but it does classify tokens into the 8 major units described above. This was designed to address the Korean tokenizer problem that has long troubled Transformer models.
Custom dictionary
A dictionary is essential for language analysis and for applying NLP to many kinds of work.
In Bareun, a single dictionary can be defined in five categories: proper nouns, compound nouns, a compound noun splitting dictionary, verbs, and adjectives.
- You can define several dictionaries in the custom dictionary API.
- During morphological analysis, you can choose which dictionary to use.
- You can change a dictionary while an analysis is running, with no interruption at all.
Highlights
Morphological tagging accuracy
With 99.6% POS tagging accuracy and 99.7% accuracy in splitting and restoring eojeol, Bareun delivers outstanding natural language processing performance. See for yourself how far it is ahead of existing commercial products and open-source tools. The measurement method is described in Why 99.6% accuracy and How accuracy is measured.
Outstanding Korean processing
By applying 115 Korean conjugation rules (14 of which are for error detection), it solves long-standing problems of Korean morphological analysis, such as morpheme restoration and eojeol splitting, which matter in an agglutinative language. Word spacing correction is also available.
Strong adaptation to spoken language
It uses a deep learning engine that understands the conjugation principles of Korean, an agglutinative language,
so it correctly analyzes new endings from spoken language such as 감사합니당(ㅂ니당/EF, final ending) and 알겠어용(어용/EF, final ending).
The evidence is explained in Adapting to colloquial speech and neologisms.
Stanford-style visualization
In Korean, a single syllable or eojeol can carry several morphemes. The Stanford UI style visualizes annotations such as elided and contracted morphemes and morpheme types in an easy-to-read form.
Continuous improvement of our own corpus
In deep learning, the quantity and quality of training data matter a great deal. We found and fixed errors that were left in the public Sejong corpus and Everyone's Corpus (Modu Corpus). We are also still supplementing scarce training data so that it reflects how language is actually used. For details, see Model training data and Corpus refinement.
Learning unregistered words
Word embedding is a technique for representing words numerically. For the Bareun model, we included POS hints at the training data construction stage so that the model captures word meaning well. Even words the model does not know are mapped to closely related words. How this works is covered in Handling out-of-vocabulary (OOV) words and Morphology-based word embedding.
Deployed as a ready-to-run server
It is deployed as a fast server that is ready to use right away, and it always responds quickly. Updating the server alone applies every change to every client. A large deep learning model is shared by many users at once, which improves resource efficiency.
Zero-downtime, multi-tenant custom dictionaries
- Each user can create any number of custom dictionaries.
- You can modify a custom dictionary while morphological analysis keeps running without interruption.
- You can switch between registered dictionaries from one analysis to the next.
Many programming languages supported
- Bareun supports Python, Java, JavaScript, and R out of the box.
- It provides a REST API over HTTP/1.1.
- In the browser, you can use it directly with
connect-web, without a separate proxy. Connect-RPCand gRPC are provided as the base interfaces, so developers can add support for new languages or build new kinds of API. We plan to keep expanding language support.- A plugin is also available for search engines such as Elasticsearch.
Easy installation, many distributions
- Bareun can be installed and used on Linux, macOS, and Windows.
- It is registered as a service, so it starts naturally after a reboot.
- It is also distributed as a Docker image, so you can install and run it in one step.
The story behind it
Corpus
We used the Sejong corpus and Everyone's Corpus (Modu Corpus) from the National Institute of Korean Language as the base, reviewed them closely, and fixed numerous errors. We also revised tagging rules that did not fit our own standards.
For example, we do not split 누가 into 누구+가, and we do not treat 뉘 as 누구+이.
Instead, we treat 누 as a pronoun and trained them as 누+가 and 누+이.
- 21st Century Sejong Project morphological and semantic analysis corpus: 12 million eojeol, written and spoken
- Morphological analysis corpus from the National Institute of Korean Language's Everyone's Corpus v1.0: 3 million eojeol (2 million written, 1 million spoken)
- Morphological training corpus built in-house by Baikal AI: supplements low-frequency data in the National Institute of Korean Language material
- A 100-million-eojeol corpus of news articles from 1992 to 2022
- An augmented corpus created by deliberately introducing word spacing errors
The results of our ongoing improvements to the Sejong corpus and Everyone's Corpus are included, but because of copyright we cannot release them publicly and keep them internal.
Our journey
- In 2019, we started from an NLP engine built by one of Korea's top linguists, who knows Korean deeply.
- 2019 to 2022: PhD- and master's-level researchers from leading Korean universities fixed errors in the corpus by hand and reviewed analysis errors one by one to supplement the training data. We adopted and built on the guidelines of the National Institute of Korean Language, training the model to fit the nature of Korean. We stubbornly tried to show Korean as it really is.
- 2019 to 2023: Developers with backgrounds in Korean language and literature led the project.
- In 2022, through the Korea Press Foundation's "Morphological Analyzer Development Project for News Organizations", we additionally trained on 100 million eojeol of BIGKINDS data.
- In 2022, we built our own Korean ambiguity-resolution data, refined it during the "Morphological Analyzer Development Project for News Organizations", and released it.
- In 2023, Baikal AI developed the morphological analyzer Bareun together with the Korea Press Foundation. The name carries the hope that it will help parse Korean correctly ("bareun" means correct or upright) and be used for proper language use.
- In 2023, bareun.ai opened Bareun to everyone, making it easy to access and use.
- In 2023, we released a lot of code and data on GitHub.
- In 2025, building on this strong morphological analysis engine, we launched a service that includes Korean spell checking.
- In 2026, we released Bareun spell checker 3.0. A paid policy was applied to the api.bareun.ai cloud service.
Ambiguity resolution evaluation data
People are sometimes confused by this too. Because of how Korean works, 가는 can be read as 가늘다 (thin), 가다 (go), or 갈다 (grind).
We found 12,175 cases in Korean where the surface form is the same but the analysis differs.
We reviewed and checked all of this data, which shares the same surface forms, and are releasing it.
It still contains many errors, though. We hope many people will join us in fixing them.
The construction process and scale are described in Ambiguity dataset.
Learn more
To look deeper into how Bareun works and how to use it, see the documents below.
- How analysis works: Analysis pipeline overview, Transformer model architecture, Automatic compound noun decomposition, Predicate conjugation and form restoration, How word spacing correction works
- Accuracy and comparison: Why 99.6% accuracy, Transformer vs dictionary-based, Bareun vs Mecab-ko
- Custom dictionary: The five custom dictionary types at a glance, Verifying zero-downtime instant updates
- Get started: Quick look, 3 fastest ways to try it, Development overview