Sentence Segmentation

Test set for evaluating and improving Vietnamese sentence boundary detection (sent_tokenize) in underthesea.

Problem

The current PunktSentenceTokenizer in underthesea fails on several Vietnamese-specific patterns, primarily in legal text where article titles are merged with sentence bodies without punctuation boundaries.

Current Results

Category	Total	Correct	Accuracy
title_content_merge	38	0	0.0%
repeated_title	13	0	0.0%
ellipsis	1	0	0.0%
numeric_period	1	0	0.0%
quoted_speech	1	0	0.0%
abbreviation	3	2	66.7%
article_header	20	20	100.0%
article_reference	1	1	100.0%
empty_input	1	1	100.0%
multi_sentence	20	20	100.0%
no_punctuation	1	1	100.0%
single_sentence	30	30	100.0%
TOTAL	130	75	57.7%

Key Issues

Title-content merge (38 cases, 0% accuracy): Legal article titles like "Tội trốn thuế" are followed by sentence body "Người nào thực hiện..." without punctuation. sent_tokenize fails to detect this boundary.
Repeated title (13 cases, 0% accuracy): Pattern "X X là..." where the title is repeated as the subject of a definition. E.g., "Hợp đồng mượn tài sản Hợp đồng mượn tài sản là..."
Ellipsis handling (0% accuracy): "..." mid-sentence causes incorrect split.
Numeric periods (0% accuracy): Periods in numbers like "1.500.000" can cause false boundaries.

Trained Punkt Model Results

Trained NLTK PunktTrainer on Vietnamese text from 4 sources (Wikipedia, news, books, legal documents). The trained model fixes punctuation-related issues but cannot address structural patterns (title-content merge).

Category	Total	Baseline	Trained	Change
title_content_merge	38	0	0	—
repeated_title	13	0	0	—
ellipsis	1	0	1	+1
numeric_period	1	0	1	+1
quoted_speech	1	0	0	—
abbreviation	3	2	2	—
article_header	20	20	20	—
article_reference	1	1	1	—
empty_input	1	1	1	—
multi_sentence	20	20	18	-2
no_punctuation	1	1	1	—
single_sentence	30	30	30	—
TOTAL	130	75	75	0

Improvements: ellipsis handling (+1), numeric period handling (+1) Regressions: multi_sentence (-2) — trade-off from better ellipsis handling (... followed by new sentence) and quote tokenization differences.

Conclusion: Punkt (trained or not) cannot solve title-content merge (51/130 failures = 39% of test set) because these require structural understanding beyond punctuation disambiguation. A different approach is needed for this category.

Files

test_cases.json — 130 test cases with input, expected output, category, and domain
evaluate.py — Evaluation script (--improved flag for trained model)
eval_results.json — Detailed evaluation results
train_punkt.py — Fetch data + train Punkt model
punkt_params_trained.json — Trained model parameters (672 abbreviations, 378 sentence starters, 3264 collocations)
sent_tokenize.py — Tokenizer using trained model

Test Case Format

{
  "id": "vlc-6200",
  "input": "Tội ngược đãi tù binh , hàng binh Người nào ngược đãi tù binh ...",
  "expected": [
    "Tội ngược đãi tù binh , hàng binh",
    "Người nào ngược đãi tù binh , hàng binh , thì bị phạt ..."
  ],
  "category": "title_content_merge",
  "domain": "legal",
  "issue": "Title merged with sentence body without boundary"
}

Usage

# Run baseline evaluation (underthesea)
python evaluate.py
python evaluate.py -v

# Run trained model evaluation
python evaluate.py --improved -v

# Train Punkt model from scratch
python train_punkt.py

Data Source

Test cases derived from undertheseanlp/UDD-1 across 5 domains: legal (VLC), news (UVN), Wikipedia (UVW), fiction (UVB-F), non-fiction (UVB-N).

Datasets:

undertheseanlp
/

sentence-segmentation-1