Consistent Annotation of the Vietnamese Treebank

1. Introduction

Treebanks are essential for training parsers, conducting linguistic research, and developing language technologies. For Vietnamesea language with analytic morphology, flexible word order, and extensive use of classifiersbuilding a highquality treebank is especially demanding. This document presents a compact but thorough guide for achieving consistent annotation across all layers of a Vietnamese treebank, from tokenisation to dependency relations.

2. Core Design Choices

2.1. Annotation Formalism

Two major formalisms dominate modern treebanks:

• PhraseStructure (Pennstyle) trees: useful for linguistic description, easy to visualise.
• Universal Dependencies (UD): offers crosslingual compatibility, straightforward conversion, and a welldefined set of POS tags and dependency relations.

For Vietnamese, we recommend the Universal Dependencies framework because it already includes a Vietnamese mapping (viet_ud) and encourages uniformity with other languages.

2.2. Tokenisation Strategy

Vietnamese orthography uses spaces to separate syllables, not words. A robust tokeniser must merge syllables that form a lexical word. The following rules are widely accepted:

• Compound nouns and proper names (i hc i_hc).
• Verbparticle constructions (i ri i_ri).
• Numeralclassifiernoun sequences (hai con mo hai_con_mo).
• Fixed expressions and idioms (e.g., c ch c_ch).

When a tokeniser disagrees with an annotator, the annotator should document the decision and, if necessary, add a comment in the # comment field.

2.3. PartofSpeech Tagset

The UD Vietnamese tagset contains 18 coarsegrained POS tags (e.g., NOUN, VERB, ADP) and a richer set of languagespecific finegrained tags (e.g., NOUNprop for proper nouns). Consistency can be achieved by:

• Using the official UD documentation as the primary reference.
• Creating a short decisiontree for ambiguous cases (e.g., distinguishing DET from NUM when a word can function as both).
• Maintaining a shared spreadsheet of exception words with their chosen tags.

2.4. Dependency Relations

Key relations that often cause disagreement in Vietnamese are:

• case vs. aux for classifier and aspect markers.
• conj vs. cc for coordinating conjunctions.
• obl for location/temporal adjuncts that appear before the verb.

The guiding principle is to follow the syntactic function rather than surface position. When a specific construction is unclear, refer to the UD Vietnamese guidelines or the nearest wellannotated example.

3. Annotation Workflow

3.1. Preannotation

1. Run an automatic tokenizer and POS tagger (e.g., VnCoreNLP or UPosTagger).

2. Perform a quick manual sanity check on a random 5% sample: verify token merges and POS tags.

3.2. Manual Annotation

Annotators work sentence by sentence using a treebanking tool such as UDPipe or UD Annotator:

1. Confirm tokenisation; split or merge if needed.
2. Select the appropriate POS tag from a dropdown.
3. Assign a head for each token and choose a dependency label.
4. If a sentence contains a construction not covered by the standard set, add a # newrel=... comment and discuss it with the senior annotator.

3.3. Interannotator Agreement (IAA)

Every 100 sentences are doubleannotated. Compute both unlabeled and labeled attachment scores. Target thresholds:

• UAS 93%
• LAS 90%

Disagreements are reviewed in a weekly annotation board meeting, and the resolution is recorded in a shared log.

3.4. Postprocessing

After the primary annotation cycle:

• Run a consistency script that checks for orphan nodes, cycles, and duplicate IDs.
• Validate against the UD 2.13 schema using ud-validate.
• Export the final data in CoNLLU format.

4. Typical Problem Areas and Solutions

4.1. Classifier Ambiguity

Classifiers (ci, con, chic ) may function as:

• Determinants (ci bn) label det.
• Oblique arguments (c sch bng bt) label obl with case Case=With.

Rule of thumb: if the classifier directly modifies a noun, treat it as det. If it appears after a verb expressing a means or instrument, treat it as obl.

4.2. Aspect and Modal Particles

Particles such as , ang, s, m, nh are usually annotated as aux. However, nh can also function as a discourse particle without syntactic dependence. In such cases, attach it to the predicate with discourse and set AdvType=Discourse.

4.3. Serial Verb Constructions

Example: ti i mua sch (I go buy a book). The main predicate is mua. The preceding verb i is labelled advcl with VerbForm=Fin and Aspect=Prog. This mirrors the treatment in other UD languages for serial verbs.

4.4. Proper Noun Segmentation

Names often consist of multiple tokens (H Ch Minh). Merge them into a single token during preprocessing, then assign PROPN. If a name contains a title (ng, B), keep the title separate and attach it with nmod:poss.

4.5. Coordination without Explicit Conjunction

Sentences like c v tm (fish and shrimp) contain the conjunction v and are straightforward. In contrast, c, tm, mc uses commas only. Annotate the first element as the head, the rest as conj, and attach commas as punct to the first coordinated element.

5. Quality Assurance Tools

Several opensource utilities help maintain consistency:

• UDPipe for tokenisation, POS tagging, and parsing; also provides validation scripts.
• CoNLLU validator checks compliance with the CoNLLU schema.
• TreeTagger useful for crosschecking POS assignments.
• Python scripts (e.g., check_missing_heads.py, check_duplicate_tokens.py) customisable to the project's conventions.

Integrate these tools into a continuousintegration pipeline so that each commit triggers automatic validation.

6. Documentation and Training

To ensure new annotators adopt the same standards:

1. Create a concise Annotation Handbook summarising the rules above, enriched with 30+ annotated examples.
2. Offer a twoday handson workshop that walks trainees through the tool, discusses frequent pitfalls, and includes a live annotation session.
3. Maintain an online FAQ where annotators can post edge cases and receive quick feedback.

7. Future Extensions

While the current treebank focuses on newswire and literary texts, the same guidelines can be adapted for:

• Socialmedia posts (handling emojis, hashtags, and nonstandard spellings).
• Spokenlanguage transcriptions (including filler words and disfluencies).
• Dialectal variants (adding languagespecific features such as Dial=North).

Each new genre should start with a pilot annotation phase, after which the guidelines are supplemented rather than overhauled.

8. Conclusion

Consistent annotation of a Vietnamese treebank demands clear decisions on tokenisation, POS tagging, and dependency relations, combined with rigorous validation and collaborative review. By adhering to the Universal Dependencies framework, employing systematic workflows, and providing thorough training, a highquality, reusable syntactic resource can be produced. Such a resource will boost parser performance, enable crosslingual research, and support downstream applications ranging from machine translation to information extraction.